{
  "schema": 1,
  "run_id": "essai-03-2026-11-14",
  "suite": "BPB2 v0.3-pilot",
  "cases": [
    {
      "id": "BPB2-01",
      "self": {
        "initial": {
          "reviewer": "gpt-6-astra · answering agent",
          "method": "Unverified self-assessment before feedback; common assessment/source packaging added afterward",
          "assessed_at": "2026-11-14T18:04:51Z",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 3,
            "B": 3
          },
          "reasons": {
            "C": "Cost-base allocation is mutually exclusive, cap-aware and invoice-reconciled. Capital and operating results match independently specified expected values for Open/Studio and heat/gas with/without donor. 3,480 allocation variants passed invariants. Risk allowance remains £45k.",
            "D": "Leads with a feasible conditional commission and the explicit January versus pre-Christmas tradeoff. Includes a serious Open-gas alternative, capital/service stop conditions and the recurring donor trap, with authority separated.",
            "E": "Eight source records are preserved and linked beside claims. Forecasts, calculations, proposals, hypothetical funding and unaccepted terms are distinguished; sensitivity uses no invented probability. No future acceptance is asserted.",
            "F": "Includes scope/exclusions, cost-base funding, annual operating bridge, dates, acceptance and benefit owners, short decision schedule, failure consequences, controls, reset, original evidence and gallery link.",
            "M": "Professional authorisation brief with fixed recommendation, clear navigation and responsive tables. It is relatively detailed, and visual rendering has not been independently assessed.",
            "B": "JSDOM executed the page without errors; required choices, sensitivity, donor, funding validation, roof allocation, comparison, reset, reload and anchors passed. Real-browser rendering/keyboard/accessibility journeys remain unverified."
          },
          "material_failures": [],
          "rationale": "Self-assessment only.",
          "artifact_sha256": "d61ef2074b254aa4dad98c3f2e96173e205b3d47fff55b555a3e861d35daa31a",
          "verification_gaps": []
        },
        "current": {
          "reviewer": "gpt-6-astra · answering agent",
          "method": "Unverified self-assessment before feedback; common assessment/source packaging added afterward; current case analysis unchanged",
          "assessed_at": "2026-11-14T18:04:51Z",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 3,
            "B": 3
          },
          "reasons": {
            "C": "Cost-base allocation is mutually exclusive, cap-aware and invoice-reconciled. Capital and operating results match independently specified expected values for Open/Studio and heat/gas with/without donor. 3,480 allocation variants passed invariants. Risk allowance remains £45k.",
            "D": "Leads with a feasible conditional commission and the explicit January versus pre-Christmas tradeoff. Includes a serious Open-gas alternative, capital/service stop conditions and the recurring donor trap, with authority separated.",
            "E": "Eight source records are preserved and linked beside claims. Forecasts, calculations, proposals, hypothetical funding and unaccepted terms are distinguished; sensitivity uses no invented probability. No future acceptance is asserted.",
            "F": "Includes scope/exclusions, cost-base funding, annual operating bridge, dates, acceptance and benefit owners, short decision schedule, failure consequences, controls, reset, original evidence and gallery link.",
            "M": "Professional authorisation brief with fixed recommendation, clear navigation and responsive tables. It is relatively detailed, and visual rendering has not been independently assessed.",
            "B": "JSDOM executed the page without errors; required choices, sensitivity, donor, funding validation, roof allocation, comparison, reset, reload and anchors passed. Real-browser rendering/keyboard/accessibility journeys remain unverified."
          },
          "material_failures": [],
          "rationale": "Self-assessment only.",
          "artifact_sha256": "d61ef2074b254aa4dad98c3f2e96173e205b3d47fff55b555a3e861d35daa31a",
          "verification_gaps": []
        }
      },
      "checked": {
        "initial": {
          "case_id": "BPB2-01",
          "reviewer": "gpt-6-astra · separate assessment agent",
          "method": "Separate source/logic review, independent checks and coordinator ordinary-browser observations; same model family, not a human or Claude assessment",
          "assessed_at": "2026-11-14T18:30:23.049742+00:00",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 4,
            "B": 4
          },
          "reasons": {
            "C": "Independently checked 162 layout/plant/donor/session/heat-cost/roof-allocation settings. Disjoint invoice bases reconcile; Open/heat council £227.15k, Open/gas £232k, donor Studio/heat £225.35k. Staffing and recurring gaps reconcile, including £14,584 donor Studio/heat gap and £9,784 five-session Open/heat gap. Opening and failed-grant consequences are correct.",
            "D": "A specific Open/heat, three-session, 17 January commission is prioritised. Open/gas is a serious £4.85k-more-capital/pre-Christmas alternative with an explicit switch condition. Owners, grant acceptance, staffed-service and commissioning gates distinguish ordinary scope selection from new committee/funder decisions.",
            "E": "Y1–Y8 claims are traceable. Donor/naming consent remains absent, forecasts and proposals remain unapproved, grants remain conditional, and £45k is a protected provision rather than spare service funding. Heat sensitivity carries through without invented probabilities.",
            "F": "Scope, exclusions, acceptance and benefits owners, invoice allocation, capital and recurring bridges, alternative, opening and dated stop/go schedule form one authorisable conditional commission.",
            "M": "The opening recommendation states the service, date and annual margin before supporting tables. A fixed dated proposal, labelled hypothetical explorer, decision schedule and source disclosures make detailed reconciliation available without obscuring the decision.",
            "B": "Coordinator ordinary-browser observations cover Studio/donor propagation, five-session funding, invalid negative funding and reload. Independent JSDOM interaction checks cover all six distinct layout/plant/donor combinations, displayed capital/opening propagation, invalid handling and reset. No runtime or internal-anchor errors found."
          },
          "material_failures": [],
          "findings": [],
          "verification_gaps": [
            "This is a separate AI assessment in the same model family, not a human professional check, Claude assessment or evidence of real-world project authority.",
            "Ordinary-browser evidence is the coordinator’s targeted observed journeys; this reviewer ran independent calculations and JSDOM interaction checks, not independent CUA, a full accessibility audit or every device/browser combination."
          ],
          "disposition": "verification only",
          "independent_checks": [
            "162 model settings checked for capital, disjoint bases, staff and operating position; 2,295 granular assertions pass.",
            "Baseline Open/heat £227.15k council; donor Studio/heat £14,584 recurring gap; five youth sessions £9,784 gap."
          ],
          "artifact_sha256": "d61ef2074b254aa4dad98c3f2e96173e205b3d47fff55b555a3e861d35daa31a",
          "status": "provisional AI assessment",
          "rationale": "AI review disposition: verification only",
          "served_artifact_sha256": "efc96f47e4f2944ad12efe3696f5f188d15e09ab1587e404b668c1d871239896",
          "review_evidence_id": "essai03-separate-review-01-07"
        },
        "current": {
          "case_id": "BPB2-01",
          "reviewer": "gpt-6-astra · separate assessment agent",
          "method": "Separate source/logic review, independent checks and coordinator ordinary-browser observations; same model family, not a human or Claude assessment; current case analysis is unchanged from the retained first answer",
          "assessed_at": "2026-11-14T18:30:23.049742+00:00",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 4,
            "B": 4
          },
          "reasons": {
            "C": "Independently checked 162 layout/plant/donor/session/heat-cost/roof-allocation settings. Disjoint invoice bases reconcile; Open/heat council £227.15k, Open/gas £232k, donor Studio/heat £225.35k. Staffing and recurring gaps reconcile, including £14,584 donor Studio/heat gap and £9,784 five-session Open/heat gap. Opening and failed-grant consequences are correct.",
            "D": "A specific Open/heat, three-session, 17 January commission is prioritised. Open/gas is a serious £4.85k-more-capital/pre-Christmas alternative with an explicit switch condition. Owners, grant acceptance, staffed-service and commissioning gates distinguish ordinary scope selection from new committee/funder decisions.",
            "E": "Y1–Y8 claims are traceable. Donor/naming consent remains absent, forecasts and proposals remain unapproved, grants remain conditional, and £45k is a protected provision rather than spare service funding. Heat sensitivity carries through without invented probabilities.",
            "F": "Scope, exclusions, acceptance and benefits owners, invoice allocation, capital and recurring bridges, alternative, opening and dated stop/go schedule form one authorisable conditional commission.",
            "M": "The opening recommendation states the service, date and annual margin before supporting tables. A fixed dated proposal, labelled hypothetical explorer, decision schedule and source disclosures make detailed reconciliation available without obscuring the decision.",
            "B": "Coordinator ordinary-browser observations cover Studio/donor propagation, five-session funding, invalid negative funding and reload. Independent JSDOM interaction checks cover all six distinct layout/plant/donor combinations, displayed capital/opening propagation, invalid handling and reset. No runtime or internal-anchor errors found."
          },
          "material_failures": [],
          "findings": [],
          "verification_gaps": [
            "This is a separate AI assessment in the same model family, not a human professional check, Claude assessment or evidence of real-world project authority.",
            "Ordinary-browser evidence is the coordinator’s targeted observed journeys; this reviewer ran independent calculations and JSDOM interaction checks, not independent CUA, a full accessibility audit or every device/browser combination."
          ],
          "disposition": "verification only",
          "independent_checks": [
            "162 model settings checked for capital, disjoint bases, staff and operating position; 2,295 granular assertions pass.",
            "Baseline Open/heat £227.15k council; donor Studio/heat £14,584 recurring gap; five youth sessions £9,784 gap."
          ],
          "artifact_sha256": "d61ef2074b254aa4dad98c3f2e96173e205b3d47fff55b555a3e861d35daa31a",
          "status": "provisional AI assessment",
          "rationale": "AI review disposition: verification only",
          "served_artifact_sha256": "819492eed4aa399300fbcb5fbdcf3e052e5bbd205c9c4b77166ff39eda62ca98",
          "review_evidence_id": "essai03-separate-review-01-07"
        }
      },
      "status": "first attempt retained"
    },
    {
      "id": "BPB2-02",
      "self": {
        "initial": {
          "reviewer": "gpt-6-astra · answering agent",
          "method": "Unverified self-assessment before feedback; common assessment/source packaging added afterward",
          "assessed_at": "2026-11-14T17:45:00Z",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 3,
            "B": 2
          },
          "reasons": {
            "C": "All eight specified sensitivities pass independent Python annual-cash construction and exhaustive signal-rule enumeration, including 1,355 numerical assertions. Capital, standing costs, grants, discounting, joint probabilities and payment-point liquidity are represented.",
            "D": "Names an executable initial policy and one action per observable signal; compares expected value, branch losses, grants and peak cash; shows monetary winners vetoed by funding and a quantified cautious alternative.",
            "E": "F1–F8 source text is preserved and linked; forecasts, hypothetical sensitivity settings and unapproved recommendations are separated. Within-signal future states remain unresolved and no state-specific gate choice is permitted.",
            "F": "Includes original dated proposal, all three policy structures, both flexible gate rule tables, six annual cash trails per policy, eight sensitivity cases, probability workings, funding visualisation and reset/reload behaviour.",
            "M": "Decision-first page with progressive cash/evidence detail and source IDs. Native browser layout and narrow-screen readability remain visually unverified.",
            "B": "55 JSDOM interaction checks pass, but no native browser executable is available. JSDOM does not prove real browser rendering, focus, scrolling or gallery navigation."
          },
          "material_failures": [],
          "rationale": "B is below 3 because ordinary browser verification has not occurred.",
          "artifact_sha256": "98d04a7026d0981c983fda1786239d1a89b7dd6098fceb6fbf8dbc1b3d485902",
          "verification_gaps": [
            "Native browser rendering and ordinary interface journeys remain unverified."
          ]
        },
        "current": {
          "reviewer": "gpt-6-astra · answering agent",
          "method": "Unverified self-assessment before feedback; common assessment/source packaging added afterward; current case analysis unchanged",
          "assessed_at": "2026-11-14T17:45:00Z",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 3,
            "B": 2
          },
          "reasons": {
            "C": "All eight specified sensitivities pass independent Python annual-cash construction and exhaustive signal-rule enumeration, including 1,355 numerical assertions. Capital, standing costs, grants, discounting, joint probabilities and payment-point liquidity are represented.",
            "D": "Names an executable initial policy and one action per observable signal; compares expected value, branch losses, grants and peak cash; shows monetary winners vetoed by funding and a quantified cautious alternative.",
            "E": "F1–F8 source text is preserved and linked; forecasts, hypothetical sensitivity settings and unapproved recommendations are separated. Within-signal future states remain unresolved and no state-specific gate choice is permitted.",
            "F": "Includes original dated proposal, all three policy structures, both flexible gate rule tables, six annual cash trails per policy, eight sensitivity cases, probability workings, funding visualisation and reset/reload behaviour.",
            "M": "Decision-first page with progressive cash/evidence detail and source IDs. Native browser layout and narrow-screen readability remain visually unverified.",
            "B": "55 JSDOM interaction checks pass, but no native browser executable is available. JSDOM does not prove real browser rendering, focus, scrolling or gallery navigation."
          },
          "material_failures": [],
          "rationale": "B is below 3 because ordinary browser verification has not occurred.",
          "artifact_sha256": "98d04a7026d0981c983fda1786239d1a89b7dd6098fceb6fbf8dbc1b3d485902",
          "verification_gaps": [
            "Native browser rendering and ordinary interface journeys remain unverified."
          ]
        }
      },
      "checked": {
        "initial": {
          "case_id": "BPB2-02",
          "reviewer": "gpt-6-astra · separate assessment agent",
          "method": "Separate source/logic review, independent checks and coordinator ordinary-browser observations; same model family, not a human or Claude assessment",
          "assessed_at": "2026-11-14T18:30:23.049742+00:00",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 4,
            "B": 4
          },
          "reasons": {
            "C": "Independent enumeration of F plus eight P and eight D observable signal rules matches all branch cash trails and all eight required sensitivity combinations, plus four supplementary settings. Base P/G-only expansion NPV £220.468k and peak £750k; g=.90 gives F £371.324k; £1,050k ceiling restores P. Year-end grant/capital timing and undiscounted branch draws are correct.",
            "D": "The page gives an implementable G/A/R delegation, initial owner and cash authority, the conditional G expansion value, branch losses, and a quantified cautious no-expansion alternative. Savings, mix and liquidity switch conditions are explicit rather than clairvoyant state choices.",
            "E": "Joint forecast probabilities are preserved; future states within each gate signal remain indistinguishable. Source/forecast/derived/proposed statuses and funding-versus-economics distinctions remain visible, with grants conditional and no invented finance.",
            "F": "Initial policy, gate decisions, six cash trails, expected and branch NPV, peak draw, grants, all requested re-optimised sensitivities and approval boundaries are coherently linked.",
            "M": "A short investment instruction leads to a fixed original record, live policy comparison, observable-rule explanation and expandable six-branch workings. Units, timing and sunk-cost treatment are adjacent to the figures.",
            "B": "Coordinator observed P→F→P on mix and ceiling changes and baseline reset. Independent interaction checks exercise all 12 supplied-control combinations and all three policy inspections, confirm displayed metrics, six cash trails and reset. No runtime or internal-anchor errors found."
          },
          "material_failures": [],
          "findings": [],
          "verification_gaps": [
            "This is a separate AI assessment in the same model family, not a human professional check, Claude assessment or evidence of real-world project authority.",
            "Ordinary-browser evidence is the coordinator’s targeted observed journeys; this reviewer ran independent calculations and JSDOM interaction checks, not independent CUA, a full accessibility audit or every device/browser combination."
          ],
          "disposition": "verification only",
          "independent_checks": [
            "Enumerated 17 information-feasible policies per setting independently; compared 12 settings including all eight mandatory sensitivities and six cash trails per policy.",
            "2,568 granular assertions pass; base P £220.468085k, high-G F £371.324138k, tightened ceiling P £348.212288k at high-G."
          ],
          "artifact_sha256": "98d04a7026d0981c983fda1786239d1a89b7dd6098fceb6fbf8dbc1b3d485902",
          "status": "provisional AI assessment",
          "rationale": "AI review disposition: verification only",
          "served_artifact_sha256": "1c0c3af62af2ffee34db7d8d736e7535af1202d63c8c747d1ea3bb43845fb7fa",
          "review_evidence_id": "essai03-separate-review-01-07"
        },
        "current": {
          "case_id": "BPB2-02",
          "reviewer": "gpt-6-astra · separate assessment agent",
          "method": "Separate source/logic review, independent checks and coordinator ordinary-browser observations; same model family, not a human or Claude assessment; current case analysis is unchanged from the retained first answer",
          "assessed_at": "2026-11-14T18:30:23.049742+00:00",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 4,
            "B": 4
          },
          "reasons": {
            "C": "Independent enumeration of F plus eight P and eight D observable signal rules matches all branch cash trails and all eight required sensitivity combinations, plus four supplementary settings. Base P/G-only expansion NPV £220.468k and peak £750k; g=.90 gives F £371.324k; £1,050k ceiling restores P. Year-end grant/capital timing and undiscounted branch draws are correct.",
            "D": "The page gives an implementable G/A/R delegation, initial owner and cash authority, the conditional G expansion value, branch losses, and a quantified cautious no-expansion alternative. Savings, mix and liquidity switch conditions are explicit rather than clairvoyant state choices.",
            "E": "Joint forecast probabilities are preserved; future states within each gate signal remain indistinguishable. Source/forecast/derived/proposed statuses and funding-versus-economics distinctions remain visible, with grants conditional and no invented finance.",
            "F": "Initial policy, gate decisions, six cash trails, expected and branch NPV, peak draw, grants, all requested re-optimised sensitivities and approval boundaries are coherently linked.",
            "M": "A short investment instruction leads to a fixed original record, live policy comparison, observable-rule explanation and expandable six-branch workings. Units, timing and sunk-cost treatment are adjacent to the figures.",
            "B": "Coordinator observed P→F→P on mix and ceiling changes and baseline reset. Independent interaction checks exercise all 12 supplied-control combinations and all three policy inspections, confirm displayed metrics, six cash trails and reset. No runtime or internal-anchor errors found."
          },
          "material_failures": [],
          "findings": [],
          "verification_gaps": [
            "This is a separate AI assessment in the same model family, not a human professional check, Claude assessment or evidence of real-world project authority.",
            "Ordinary-browser evidence is the coordinator’s targeted observed journeys; this reviewer ran independent calculations and JSDOM interaction checks, not independent CUA, a full accessibility audit or every device/browser combination."
          ],
          "disposition": "verification only",
          "independent_checks": [
            "Enumerated 17 information-feasible policies per setting independently; compared 12 settings including all eight mandatory sensitivities and six cash trails per policy.",
            "2,568 granular assertions pass; base P £220.468085k, high-G F £371.324138k, tightened ceiling P £348.212288k at high-G."
          ],
          "artifact_sha256": "98d04a7026d0981c983fda1786239d1a89b7dd6098fceb6fbf8dbc1b3d485902",
          "status": "provisional AI assessment",
          "rationale": "AI review disposition: verification only",
          "served_artifact_sha256": "39829590366bc605ebc53e1bfd952b7aaf93f13a8fc6415d18dd140618c7498f",
          "review_evidence_id": "essai03-separate-review-01-07"
        }
      },
      "status": "first attempt retained"
    },
    {
      "id": "BPB2-03",
      "self": {
        "initial": {
          "reviewer": "gpt-6-astra · answering agent",
          "method": "Unverified self-assessment before feedback; common assessment/source packaging added afterward",
          "assessed_at": "2026-11-14T17:44:59Z",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 3,
            "B": 3
          },
          "reasons": {
            "C": "Enumerates all32 offer combinations and14,400 assignment/resource-order cases. Each retained schedule separately satisfies all18 activities, durations, precedences, calendar days, whole-crew assignment, fixed Q slots and resource non-overlap. Cheapest S18 and original opening37 verified by exhaustive comparison.",
            "D": "Clear S + T24 recommendation, cheaper S18 fallback, no-purchase37 and £14k holding comparison. Shows resource sequences, P-to-Q margin, sponsor/board authority and decisions still required. Quantifies conditional extra-buyout break-even without inventing probability.",
            "E": "M1–M8 remain recoverable with dates and source anchors. Distinguishes proposed schedules and trial changes from approval, booking and acceptance evidence. Deterministic optimality scope and lack of duration-risk data are explicit.",
            "F": "All32 purchase combinations select full feasible schedules; all18 activities, resource lanes and waiting reasons are shown. Manual moves detect invalid assignments/calendar/precedence/overlap and withdraw opening; Q/R move to next booked slot when P slips. Reset and reload restore recommendation.",
            "M": "Practical recommendation and three-option comparison lead. Tables are auditable and sources remain accessible; the complete18-activity plan and32-option comparison are necessarily detailed, and actual visual rendering was not inspected.",
            "B": "Sixteen JSDOM journeys pass with no script errors, including all32 combinations via actual controls, invalid inputs, calendar/crew conflicts, reslotting, reset and reload. Native Chromium launch was attempted but executable unavailable; no real browser or responsive rendering claim is made."
          },
          "material_failures": [],
          "rationale": "Self-assessed readiness only; native-browser rendering remains unverified",
          "artifact_sha256": "2dabd75026df55f7a0fdfb4ea1c864ece9d936096696f71a5d503af6601d045c",
          "verification_gaps": [
            "Native browser rendering and assistive-technology behaviour not tested; coordinator ordinary-browser check remains separate.",
            "Manual activity trials hold other starts fixed except Q/R; they validate rather than automatically re-level the full plan.",
            "No stochastic risk optimisation, alternative future offers, external certification or live reservations."
          ]
        },
        "current": {
          "reviewer": "gpt-6-astra · answering agent",
          "method": "Unverified self-assessment before feedback; common assessment/source packaging added afterward; current case analysis unchanged",
          "assessed_at": "2026-11-14T17:44:59Z",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 3,
            "B": 3
          },
          "reasons": {
            "C": "Enumerates all32 offer combinations and14,400 assignment/resource-order cases. Each retained schedule separately satisfies all18 activities, durations, precedences, calendar days, whole-crew assignment, fixed Q slots and resource non-overlap. Cheapest S18 and original opening37 verified by exhaustive comparison.",
            "D": "Clear S + T24 recommendation, cheaper S18 fallback, no-purchase37 and £14k holding comparison. Shows resource sequences, P-to-Q margin, sponsor/board authority and decisions still required. Quantifies conditional extra-buyout break-even without inventing probability.",
            "E": "M1–M8 remain recoverable with dates and source anchors. Distinguishes proposed schedules and trial changes from approval, booking and acceptance evidence. Deterministic optimality scope and lack of duration-risk data are explicit.",
            "F": "All32 purchase combinations select full feasible schedules; all18 activities, resource lanes and waiting reasons are shown. Manual moves detect invalid assignments/calendar/precedence/overlap and withdraw opening; Q/R move to next booked slot when P slips. Reset and reload restore recommendation.",
            "M": "Practical recommendation and three-option comparison lead. Tables are auditable and sources remain accessible; the complete18-activity plan and32-option comparison are necessarily detailed, and actual visual rendering was not inspected.",
            "B": "Sixteen JSDOM journeys pass with no script errors, including all32 combinations via actual controls, invalid inputs, calendar/crew conflicts, reslotting, reset and reload. Native Chromium launch was attempted but executable unavailable; no real browser or responsive rendering claim is made."
          },
          "material_failures": [],
          "rationale": "Self-assessed readiness only; native-browser rendering remains unverified",
          "artifact_sha256": "2dabd75026df55f7a0fdfb4ea1c864ece9d936096696f71a5d503af6601d045c",
          "verification_gaps": [
            "Native browser rendering and assistive-technology behaviour not tested; coordinator ordinary-browser check remains separate.",
            "Manual activity trials hold other starts fixed except Q/R; they validate rather than automatically re-level the full plan.",
            "No stochastic risk optimisation, alternative future offers, external certification or live reservations."
          ]
        }
      },
      "checked": {
        "initial": {
          "case_id": "BPB2-03",
          "reviewer": "gpt-6-astra · separate assessment agent",
          "method": "Separate source/logic review, independent checks and coordinator ordinary-browser observations; same model family, not a human or Claude assessment",
          "assessed_at": "2026-11-14T18:30:23.049742+00:00",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 4,
            "B": 4
          },
          "reasons": {
            "C": "Independently rebuilt the resource-order enumeration from M1–M5 and audited all 32 complete 18-activity schedules. All earliest P finishes and openings match: original P29/Q34/open37; S P27/open30; ST P24/open30; XC P27/open30. Calendars, non-preemption, whole-resource assignment and joint C0/T requirements pass.",
            "D": "The £24k ST recommendation explicitly buys three readiness boundaries for £6k over the £18k least-cost on-time S plan. It distinguishes public-date commitment from cheaper £14k delay, names sponsor/board decisions and booking commitments, and declines to invent a delay probability.",
            "E": "M7 is corrected as an unsupported overlapping-resource draft; source calendars and accepted windows govern. Calculated feasibility is separated from successful tests, approval and reservation. The optional 42.9-point break-even illustration is expressly conditional, not an inferred probability.",
            "F": "Every claimed combination has all activities, dependencies, resources, waiting, lanes, P-to-Q margin, purchase/holding costs and authority. Original, least-cost, recommended and resilience alternatives remain inspectable; optimality is bounded to the fixture.",
            "M": "The £24k decision, £18k alternative and baseline delay are prioritised, with selectable complete schedules and resource lanes supporting them. Boundary-date conventions and the non-probabilistic meaning of margin prevent likely misreadings.",
            "B": "Coordinator observed original/S/ST selection, a P move propagating to later acceptance/opening, and reload. Independent interactions exercise 32 package selections, an unavailable S0 assignment, a P move, invalid negative start and reset. Critical opening claims are withdrawn for conflicts. Local support-link packaging omission was coordinator-side and corrected, not a model repair."
          },
          "material_failures": [],
          "findings": [
            {
              "severity": "observation",
              "description": "Coordinator reported an initial serving/package omission of four otherwise-valid local support-file links, then restored unchanged support files and relocated links.",
              "evidence": "Coordinator message received during review; raw answer contains the support links. Their self-score target was not read.",
              "proposed_repair": "No substantive model repair. Retain the coordinator packaging note separately from first-answer PM/browser evaluation."
            }
          ],
          "verification_gaps": [
            "This is a separate AI assessment in the same model family, not a human professional check, Claude assessment or evidence of real-world project authority.",
            "Ordinary-browser evidence is the coordinator’s targeted observed journeys; this reviewer ran independent calculations and JSDOM interaction checks, not independent CUA, a full accessibility audit or every device/browser combination."
          ],
          "disposition": "verification only",
          "independent_checks": [
            "Independent complete assignment/resource-order enumeration for all 32 purchase combinations and per-day witness audits; 4,736 granular assertions pass.",
            "Original P29/open37; S P27/open30; ST P24/open30; XC P27/open30. Invalid UI move withdraws opening and P26 moves Q34/open37."
          ],
          "artifact_sha256": "2dabd75026df55f7a0fdfb4ea1c864ece9d936096696f71a5d503af6601d045c",
          "status": "provisional AI assessment",
          "rationale": "AI review disposition: verification only",
          "served_artifact_sha256": "11d88704ec6df655af09d9adf8d952057dea3c6be1ff551d9855706ebebfa889",
          "review_evidence_id": "essai03-separate-review-01-07"
        },
        "current": {
          "case_id": "BPB2-03",
          "reviewer": "gpt-6-astra · separate assessment agent",
          "method": "Separate source/logic review, independent checks and coordinator ordinary-browser observations; same model family, not a human or Claude assessment; current case analysis is unchanged from the retained first answer",
          "assessed_at": "2026-11-14T18:30:23.049742+00:00",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 4,
            "B": 4
          },
          "reasons": {
            "C": "Independently rebuilt the resource-order enumeration from M1–M5 and audited all 32 complete 18-activity schedules. All earliest P finishes and openings match: original P29/Q34/open37; S P27/open30; ST P24/open30; XC P27/open30. Calendars, non-preemption, whole-resource assignment and joint C0/T requirements pass.",
            "D": "The £24k ST recommendation explicitly buys three readiness boundaries for £6k over the £18k least-cost on-time S plan. It distinguishes public-date commitment from cheaper £14k delay, names sponsor/board decisions and booking commitments, and declines to invent a delay probability.",
            "E": "M7 is corrected as an unsupported overlapping-resource draft; source calendars and accepted windows govern. Calculated feasibility is separated from successful tests, approval and reservation. The optional 42.9-point break-even illustration is expressly conditional, not an inferred probability.",
            "F": "Every claimed combination has all activities, dependencies, resources, waiting, lanes, P-to-Q margin, purchase/holding costs and authority. Original, least-cost, recommended and resilience alternatives remain inspectable; optimality is bounded to the fixture.",
            "M": "The £24k decision, £18k alternative and baseline delay are prioritised, with selectable complete schedules and resource lanes supporting them. Boundary-date conventions and the non-probabilistic meaning of margin prevent likely misreadings.",
            "B": "Coordinator observed original/S/ST selection, a P move propagating to later acceptance/opening, and reload. Independent interactions exercise 32 package selections, an unavailable S0 assignment, a P move, invalid negative start and reset. Critical opening claims are withdrawn for conflicts. Local support-link packaging omission was coordinator-side and corrected, not a model repair."
          },
          "material_failures": [],
          "findings": [
            {
              "severity": "observation",
              "description": "Coordinator reported an initial serving/package omission of four otherwise-valid local support-file links, then restored unchanged support files and relocated links.",
              "evidence": "Coordinator message received during review; raw answer contains the support links. Their self-score target was not read.",
              "proposed_repair": "No substantive model repair. Retain the coordinator packaging note separately from first-answer PM/browser evaluation."
            }
          ],
          "verification_gaps": [
            "This is a separate AI assessment in the same model family, not a human professional check, Claude assessment or evidence of real-world project authority.",
            "Ordinary-browser evidence is the coordinator’s targeted observed journeys; this reviewer ran independent calculations and JSDOM interaction checks, not independent CUA, a full accessibility audit or every device/browser combination."
          ],
          "disposition": "verification only",
          "independent_checks": [
            "Independent complete assignment/resource-order enumeration for all 32 purchase combinations and per-day witness audits; 4,736 granular assertions pass.",
            "Original P29/open37; S P27/open30; ST P24/open30; XC P27/open30. Invalid UI move withdraws opening and P26 moves Q34/open37."
          ],
          "artifact_sha256": "2dabd75026df55f7a0fdfb4ea1c864ece9d936096696f71a5d503af6601d045c",
          "status": "provisional AI assessment",
          "rationale": "AI review disposition: verification only",
          "served_artifact_sha256": "91feaf9c51876af5115f4b3cb82a656348e651509efff6095f4c96368fe1850c",
          "review_evidence_id": "essai03-separate-review-01-07"
        }
      },
      "status": "first attempt retained"
    },
    {
      "id": "BPB2-04",
      "self": {
        "initial": {
          "reviewer": "gpt-6-astra · answering agent",
          "method": "Unverified self-assessment before feedback; common assessment/source packaging added afterward",
          "assessed_at": "2026-11-14T18:03:00Z",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 3,
            "B": 3
          },
          "reasons": {
            "C": "Cost, accrual, advance, retention, inventory and PO bridges reconcile; baseline and 18 quantity/stock combinations pass conservation checks. Current provision replaces old allowance.",
            "D": "Leads with board funding recommendation and timing; distinguishes package revisions, recognised liabilities and spend authority; assigns a small set of control actions.",
            "E": "Material claims cite B1–B8; complete source retained; unresolved cable stays provisional; changes are explicitly hypothetical; no probability or approval invented.",
            "F": "Includes package EAC, recognised cost, future cost, cash by completion and later, inventory balances, duplicate/superseded/cancelled commitments, working scenario controls, reset and gallery route.",
            "M": "Clear practitioner-led recommendation with section navigation and inspectable tables. The detailed reconciliation is lengthy and wide tables require horizontal scrolling on small screens.",
            "B": "404 local JSDOM assertions passed, covering inputs, outputs, error recovery, reset/reload, source anchors and required route. Real-browser rendering and ordinary browser journeys remain unverified."
          },
          "material_failures": [],
          "rationale": "Self-assessment only.",
          "artifact_sha256": "c13dfda819de7f19f23e7b802d661a4617fa17fcbd4bc0f0adf92ce7769cb842",
          "verification_gaps": []
        },
        "current": {
          "reviewer": "gpt-6-astra · answering agent",
          "method": "Unverified self-assessment before feedback; common assessment/source packaging added afterward; current case analysis unchanged",
          "assessed_at": "2026-11-14T18:03:00Z",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 3,
            "B": 3
          },
          "reasons": {
            "C": "Cost, accrual, advance, retention, inventory and PO bridges reconcile; baseline and 18 quantity/stock combinations pass conservation checks. Current provision replaces old allowance.",
            "D": "Leads with board funding recommendation and timing; distinguishes package revisions, recognised liabilities and spend authority; assigns a small set of control actions.",
            "E": "Material claims cite B1–B8; complete source retained; unresolved cable stays provisional; changes are explicitly hypothetical; no probability or approval invented.",
            "F": "Includes package EAC, recognised cost, future cost, cash by completion and later, inventory balances, duplicate/superseded/cancelled commitments, working scenario controls, reset and gallery route.",
            "M": "Clear practitioner-led recommendation with section navigation and inspectable tables. The detailed reconciliation is lengthy and wide tables require horizontal scrolling on small screens.",
            "B": "404 local JSDOM assertions passed, covering inputs, outputs, error recovery, reset/reload, source anchors and required route. Real-browser rendering and ordinary browser journeys remain unverified."
          },
          "material_failures": [],
          "rationale": "Self-assessment only.",
          "artifact_sha256": "c13dfda819de7f19f23e7b802d661a4617fa17fcbd4bc0f0adf92ce7769cb842",
          "verification_gaps": []
        }
      },
      "checked": {
        "initial": {
          "case_id": "BPB2-04",
          "reviewer": "gpt-6-astra · separate assessment agent",
          "method": "Separate source/logic review, independent checks and coordinator ordinary-browser observations; same model family, not a human or Claude assessment",
          "assessed_at": "2026-11-14T18:30:23.049742+00:00",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 4,
            "B": 4
          },
          "reasons": {
            "C": "Independent package/accrual/advance/cash checks reproduce £1,033k recognised, £1,669k EAC, £1,734k funding, £34k gap, £825.65k December cash and £19.95k later retention. Ten rock/cable settings reconcile, including found cable reducing purchases/EAC/cash £8k while received cost stays fixed.",
            "D": "The board request prioritises the whole-life increase and later retained cash. Asha/finance/purchasing, stores and engineer have specific corrections and evidence duties before commitment; order revisions remain distinct from funding authority and no invented PO approver is named.",
            "E": "Invoices, accepted/unrecorded obligations, advances, retention and original export are separately traceable to B1–B8. Unverified stock is provisionally unavailable, not declared lost; no unsupplied contractor headline is invented. The old reserve is replaced and the new provision is neither expense nor a percentile.",
            "F": "Package EAC, inventory balances, claim correction, PO bridge, cutoff liabilities, advance assets, future cash and retained tail form a complete cost-to-funding reconciliation and control recommendation.",
            "M": "The £34k decision leads, followed by cost, cash, stock and commitment sections. The fixed dated baseline remains visible beside scenario changes; units and the distinction between lifetime gap and completion cash are clear.",
            "B": "Coordinator observed combined found-cable/2,200m rock propagation, invalid 4,001m rejection and reset/reload. Independent interaction checks cover ten boundary/intermediate settings and verify displayed EAC/cash, retained valid input after rejection and baseline restore. No runtime or internal-anchor errors found."
          },
          "material_failures": [],
          "findings": [],
          "verification_gaps": [
            "This is a separate AI assessment in the same model family, not a human professional check, Claude assessment or evidence of real-world project authority.",
            "Ordinary-browser evidence is the coordinator’s targeted observed journeys; this reviewer ran independent calculations and JSDOM interaction checks, not independent CUA, a full accessibility audit or every device/browser combination."
          ],
          "disposition": "verification only",
          "independent_checks": [
            "Ten stock/rock scenarios; 100 assertions pass, including received-cost invariance, advance assets, liabilities and lifetime cash identity.",
            "Baseline recognised £1,033k; EAC£1,669k; funding£1,734k; future cash£825.65k plus£19.95k retention."
          ],
          "artifact_sha256": "c13dfda819de7f19f23e7b802d661a4617fa17fcbd4bc0f0adf92ce7769cb842",
          "status": "provisional AI assessment",
          "rationale": "AI review disposition: verification only",
          "served_artifact_sha256": "e17ed77df9b51b254f70c6a3d233f0c7a47753830a544ecc0de5fbea98e087ee",
          "review_evidence_id": "essai03-separate-review-01-07"
        },
        "current": {
          "case_id": "BPB2-04",
          "reviewer": "gpt-6-astra · separate assessment agent",
          "method": "Separate source/logic review, independent checks and coordinator ordinary-browser observations; same model family, not a human or Claude assessment; current case analysis is unchanged from the retained first answer",
          "assessed_at": "2026-11-14T18:30:23.049742+00:00",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 4,
            "B": 4
          },
          "reasons": {
            "C": "Independent package/accrual/advance/cash checks reproduce £1,033k recognised, £1,669k EAC, £1,734k funding, £34k gap, £825.65k December cash and £19.95k later retention. Ten rock/cable settings reconcile, including found cable reducing purchases/EAC/cash £8k while received cost stays fixed.",
            "D": "The board request prioritises the whole-life increase and later retained cash. Asha/finance/purchasing, stores and engineer have specific corrections and evidence duties before commitment; order revisions remain distinct from funding authority and no invented PO approver is named.",
            "E": "Invoices, accepted/unrecorded obligations, advances, retention and original export are separately traceable to B1–B8. Unverified stock is provisionally unavailable, not declared lost; no unsupplied contractor headline is invented. The old reserve is replaced and the new provision is neither expense nor a percentile.",
            "F": "Package EAC, inventory balances, claim correction, PO bridge, cutoff liabilities, advance assets, future cash and retained tail form a complete cost-to-funding reconciliation and control recommendation.",
            "M": "The £34k decision leads, followed by cost, cash, stock and commitment sections. The fixed dated baseline remains visible beside scenario changes; units and the distinction between lifetime gap and completion cash are clear.",
            "B": "Coordinator observed combined found-cable/2,200m rock propagation, invalid 4,001m rejection and reset/reload. Independent interaction checks cover ten boundary/intermediate settings and verify displayed EAC/cash, retained valid input after rejection and baseline restore. No runtime or internal-anchor errors found."
          },
          "material_failures": [],
          "findings": [],
          "verification_gaps": [
            "This is a separate AI assessment in the same model family, not a human professional check, Claude assessment or evidence of real-world project authority.",
            "Ordinary-browser evidence is the coordinator’s targeted observed journeys; this reviewer ran independent calculations and JSDOM interaction checks, not independent CUA, a full accessibility audit or every device/browser combination."
          ],
          "disposition": "verification only",
          "independent_checks": [
            "Ten stock/rock scenarios; 100 assertions pass, including received-cost invariance, advance assets, liabilities and lifetime cash identity.",
            "Baseline recognised £1,033k; EAC£1,669k; funding£1,734k; future cash£825.65k plus£19.95k retention."
          ],
          "artifact_sha256": "c13dfda819de7f19f23e7b802d661a4617fa17fcbd4bc0f0adf92ce7769cb842",
          "status": "provisional AI assessment",
          "rationale": "AI review disposition: verification only",
          "served_artifact_sha256": "684f30b6917d32ef92d252458db04f52c1d71c6d9de1e28f9a3267742b084669",
          "review_evidence_id": "essai03-separate-review-01-07"
        }
      },
      "status": "first attempt retained"
    },
    {
      "id": "BPB2-05",
      "self": {
        "initial": {
          "reviewer": "gpt-6-astra · answering agent",
          "method": "Unverified self-assessment before feedback; common assessment/source packaging added afterward",
          "assessed_at": "2026-11-14T18:06:03Z",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 3,
            "B": 3
          },
          "reasons": {
            "C": "All 20 policies at both survey sensitivities match separately implemented Python enumeration; costs include finite windows, remobilisation, reservation, survey and support. Whole-policy P90 and deadline constraints are applied.",
            "D": "Recommends unconditional B with its £10.58k assurance premium, compares the useful frontier and all feasible alternatives, and gives an actionable dated authorisation/call/funding sequence.",
            "E": "Original O1–O8 records are readable; source IDs accompany material claims. Forecasts, recommendations, hypothetical results and approvals are distinct. Joint states, posteriors, false results and residual unfunded tail remain explicit.",
            "F": "All four fixed and sixteen contingent policies, branch probabilities, state transfers, dates, cost components, aggregated CDF, sensitivity, funding reserve and tail metrics are included with working controls/reset.",
            "M": "Audience-specific recommendation and useful decision sections lead; detailed tables are collapsible or scrollable. The complete evidence and 20-policy comparison make this a substantial page, and real rendered readability is not yet verified.",
            "B": "Eleven local JSDOM tests passed, including every supported policy/sensitivity control combination, inspection/reset controls, evidence anchors and no script errors. No actual browser rendering was available; ordinary-interface checks remain for the coordinator."
          },
          "material_failures": [],
          "rationale": "Self-assessment only.",
          "artifact_sha256": "42719cb599272cccc1322aced06fd0322eadca8dd6be4087574630089a99de4b",
          "verification_gaps": []
        },
        "current": {
          "reviewer": "gpt-6-astra · answering agent",
          "method": "Unverified self-assessment before feedback; common assessment/source packaging added afterward; current case analysis unchanged",
          "assessed_at": "2026-11-14T18:06:03Z",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 3,
            "B": 3
          },
          "reasons": {
            "C": "All 20 policies at both survey sensitivities match separately implemented Python enumeration; costs include finite windows, remobilisation, reservation, survey and support. Whole-policy P90 and deadline constraints are applied.",
            "D": "Recommends unconditional B with its £10.58k assurance premium, compares the useful frontier and all feasible alternatives, and gives an actionable dated authorisation/call/funding sequence.",
            "E": "Original O1–O8 records are readable; source IDs accompany material claims. Forecasts, recommendations, hypothetical results and approvals are distinct. Joint states, posteriors, false results and residual unfunded tail remain explicit.",
            "F": "All four fixed and sixteen contingent policies, branch probabilities, state transfers, dates, cost components, aggregated CDF, sensitivity, funding reserve and tail metrics are included with working controls/reset.",
            "M": "Audience-specific recommendation and useful decision sections lead; detailed tables are collapsible or scrollable. The complete evidence and 20-policy comparison make this a substantial page, and real rendered readability is not yet verified.",
            "B": "Eleven local JSDOM tests passed, including every supported policy/sensitivity control combination, inspection/reset controls, evidence anchors and no script errors. No actual browser rendering was available; ordinary-interface checks remain for the coordinator."
          },
          "material_failures": [],
          "rationale": "Self-assessment only.",
          "artifact_sha256": "42719cb599272cccc1322aced06fd0322eadca8dd6be4087574630089a99de4b",
          "verification_gaps": []
        }
      },
      "checked": {
        "initial": {
          "case_id": "BPB2-05",
          "reviewer": "gpt-6-astra · separate assessment agent",
          "method": "Separate source/logic review, independent checks and coordinator ordinary-browser observations; same model family, not a human or Claude assessment",
          "assessed_at": "2026-11-14T18:30:23.049742+00:00",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 4,
            "B": 4
          },
          "reasons": {
            "C": "Independent exact enumeration matches all four fixed and 16 survey policies at both sensitivities, including every outcome cost, finish, probability, P90 and tail. B is £606k expected/£822k P90/100% model on-time. +B/−N is £595.42k/£847k/97%, with £278k excess at 3%; quality .65 lowers on-time to94.75%.",
            "D": "Unconditional B is defended by the quantified £10.58k assurance premium and lower request/tail. The cheaper compliant survey alternative is retained with explicit residual funding and deadline consequences; order/package deadlines and board/marine/inspection roles are actionable.",
            "E": "The joint prior and imperfect result information are retained, negative is not certification, and no real-world safety assurance is claimed. Planning estimates, proposed policy, hypothetical controls and no observed survey result remain distinct.",
            "F": "Policy frontier, conditional branches, vessel-window transfers, cost components, cumulative distribution, reserve/overrun, sensitivity, decision timeline and residual sponsor funding decision form a complete brief.",
            "M": "The recommended B decision and premium are concise; branch/outcome details and the full frontier are separately inspectable. Model-only 100% language, P90 versus full funding and false-negative consequences are close to the claims.",
            "B": "Coordinator observed .65 sensitivity breach, B and W switches and reload. Independent interactions exercise every fixed/survey policy at both qualities and verify displayed expected cost/P90 plus reset. Source anchors resolve and no script errors occur."
          },
          "material_failures": [],
          "findings": [],
          "verification_gaps": [
            "This is a separate AI assessment in the same model family, not a human professional check, Claude assessment or evidence of real-world project authority.",
            "Ordinary-browser evidence is the coordinator’s targeted observed journeys; this reviewer ran independent calculations and JSDOM interaction checks, not independent CUA, a full accessibility audit or every device/browser combination."
          ],
          "disposition": "verification only",
          "independent_checks": [
            "40 exact policy/quality evaluations and every joint outcome compared; 1,184 granular assertions pass.",
            "B£606k/£822k/100%; +B/−N£595.42k/£847k/97%; false-negative tail3%, max excess£278k; .65quality deadline94.75%."
          ],
          "artifact_sha256": "42719cb599272cccc1322aced06fd0322eadca8dd6be4087574630089a99de4b",
          "status": "provisional AI assessment",
          "rationale": "AI review disposition: verification only",
          "served_artifact_sha256": "9b7febb319128527bfb71764174004fe6958c71eb4efbf81cbbf67b483e6a8d4",
          "review_evidence_id": "essai03-separate-review-01-07"
        },
        "current": {
          "case_id": "BPB2-05",
          "reviewer": "gpt-6-astra · separate assessment agent",
          "method": "Separate source/logic review, independent checks and coordinator ordinary-browser observations; same model family, not a human or Claude assessment; current case analysis is unchanged from the retained first answer",
          "assessed_at": "2026-11-14T18:30:23.049742+00:00",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 4,
            "B": 4
          },
          "reasons": {
            "C": "Independent exact enumeration matches all four fixed and 16 survey policies at both sensitivities, including every outcome cost, finish, probability, P90 and tail. B is £606k expected/£822k P90/100% model on-time. +B/−N is £595.42k/£847k/97%, with £278k excess at 3%; quality .65 lowers on-time to94.75%.",
            "D": "Unconditional B is defended by the quantified £10.58k assurance premium and lower request/tail. The cheaper compliant survey alternative is retained with explicit residual funding and deadline consequences; order/package deadlines and board/marine/inspection roles are actionable.",
            "E": "The joint prior and imperfect result information are retained, negative is not certification, and no real-world safety assurance is claimed. Planning estimates, proposed policy, hypothetical controls and no observed survey result remain distinct.",
            "F": "Policy frontier, conditional branches, vessel-window transfers, cost components, cumulative distribution, reserve/overrun, sensitivity, decision timeline and residual sponsor funding decision form a complete brief.",
            "M": "The recommended B decision and premium are concise; branch/outcome details and the full frontier are separately inspectable. Model-only 100% language, P90 versus full funding and false-negative consequences are close to the claims.",
            "B": "Coordinator observed .65 sensitivity breach, B and W switches and reload. Independent interactions exercise every fixed/survey policy at both qualities and verify displayed expected cost/P90 plus reset. Source anchors resolve and no script errors occur."
          },
          "material_failures": [],
          "findings": [],
          "verification_gaps": [
            "This is a separate AI assessment in the same model family, not a human professional check, Claude assessment or evidence of real-world project authority.",
            "Ordinary-browser evidence is the coordinator’s targeted observed journeys; this reviewer ran independent calculations and JSDOM interaction checks, not independent CUA, a full accessibility audit or every device/browser combination."
          ],
          "disposition": "verification only",
          "independent_checks": [
            "40 exact policy/quality evaluations and every joint outcome compared; 1,184 granular assertions pass.",
            "B£606k/£822k/100%; +B/−N£595.42k/£847k/97%; false-negative tail3%, max excess£278k; .65quality deadline94.75%."
          ],
          "artifact_sha256": "42719cb599272cccc1322aced06fd0322eadca8dd6be4087574630089a99de4b",
          "status": "provisional AI assessment",
          "rationale": "AI review disposition: verification only",
          "served_artifact_sha256": "0779d6e2732e9f2d9381f24df61b5536469dd8e8a694b28288be0d5a02aec6cb",
          "review_evidence_id": "essai03-separate-review-01-07"
        }
      },
      "status": "first attempt retained"
    },
    {
      "id": "BPB2-06",
      "self": {
        "initial": {
          "reviewer": "gpt-6-astra · answering agent",
          "method": "Unverified self-assessment before feedback; common assessment/source packaging added afterward",
          "assessed_at": "2026-11-14T18:17:30Z",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 3,
            "B": 3
          },
          "reasons": {
            "C": "Separate Python arithmetic and JavaScript model agree across all six demand/reference cases. Annual submitted-order allowances, replacement staffing, exit, capacity windows, sequential owner sign-off, ceiling exclusions and the published scoring denominator are preserved.",
            "D": "Leads with Birch on the approved central basis, identifies the finance director as signer, makes the 19 April deadline and required purchases explicit, and supplies conditional growth-Cedar and low-Alder fallback decisions. Controls recalculate gating and ranking.",
            "E": "All eight original records are embedded with source links; offered capability, directly checked reference evidence, provisional points and future acceptance are distinguished. Confirmation remains hypothetical, and no demand probabilities or risk values are invented.",
            "F": "Includes a common cost bridge, annual charging workings, all three weekly allocations, six demand/reference combinations, purchase-failure tests, reset, original recommendation, signer, deadline, fallback and gallery route.",
            "M": "Recommendation-first professional brief, cost composition chart, labelled tables and source navigation are present. The complete supporting record makes the page long; visual and small-screen rendering have not been checked in a real browser.",
            "B": "Sixteen local JSDOM checks passed, including control interactions, all 48 staffing/demand/reference combinations, plan selection, source opening, reset and fresh-DOM reload. This is DOM evidence only; real-browser layout, scrolling and keyboard behaviour remain unverified."
          },
          "material_failures": [],
          "rationale": "Self-assessment only.",
          "artifact_sha256": "7edca985ec1f55a13a9d2b7393e75ff24757bcdf61175a0b3e6ec09eedae6e75",
          "verification_gaps": [
            "No local real-browser executable was verified; the solver did not use CUA or claim rendered-browser validation.",
            "Inputs and reservations are deliberately ephemeral hypothetical settings, not a saved procurement record or actual approval.",
            "Future Merefield acceptance and Cedar's missing customer confirmation cannot be established from the supplied evidence."
          ]
        },
        "current": {
          "reviewer": "gpt-6-astra · answering agent",
          "method": "Unverified self-assessment before feedback; common assessment/source packaging added afterward; current case analysis unchanged",
          "assessed_at": "2026-11-14T18:17:30Z",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 3,
            "B": 3
          },
          "reasons": {
            "C": "Separate Python arithmetic and JavaScript model agree across all six demand/reference cases. Annual submitted-order allowances, replacement staffing, exit, capacity windows, sequential owner sign-off, ceiling exclusions and the published scoring denominator are preserved.",
            "D": "Leads with Birch on the approved central basis, identifies the finance director as signer, makes the 19 April deadline and required purchases explicit, and supplies conditional growth-Cedar and low-Alder fallback decisions. Controls recalculate gating and ranking.",
            "E": "All eight original records are embedded with source links; offered capability, directly checked reference evidence, provisional points and future acceptance are distinguished. Confirmation remains hypothetical, and no demand probabilities or risk values are invented.",
            "F": "Includes a common cost bridge, annual charging workings, all three weekly allocations, six demand/reference combinations, purchase-failure tests, reset, original recommendation, signer, deadline, fallback and gallery route.",
            "M": "Recommendation-first professional brief, cost composition chart, labelled tables and source navigation are present. The complete supporting record makes the page long; visual and small-screen rendering have not been checked in a real browser.",
            "B": "Sixteen local JSDOM checks passed, including control interactions, all 48 staffing/demand/reference combinations, plan selection, source opening, reset and fresh-DOM reload. This is DOM evidence only; real-browser layout, scrolling and keyboard behaviour remain unverified."
          },
          "material_failures": [],
          "rationale": "Self-assessment only.",
          "artifact_sha256": "7edca985ec1f55a13a9d2b7393e75ff24757bcdf61175a0b3e6ec09eedae6e75",
          "verification_gaps": [
            "No local real-browser executable was verified; the solver did not use CUA or claim rendered-browser validation.",
            "Inputs and reservations are deliberately ephemeral hypothetical settings, not a saved procurement record or actual approval.",
            "Future Merefield acceptance and Cedar's missing customer confirmation cannot be established from the supplied evidence."
          ]
        }
      },
      "checked": {
        "initial": {
          "case_id": "BPB2-06",
          "reviewer": "gpt-6-astra · separate assessment agent",
          "method": "Separate source/logic review, independent checks and coordinator ordinary-browser observations; same model family, not a human or Claude assessment",
          "assessed_at": "2026-11-14T18:30:23.049742+00:00",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 4,
            "B": 4
          },
          "reasons": {
            "C": "Independently normalised all three demand paths and both reference states, including all eight specialist-purchase toggles. Alder £148.3/180.3/212.3k, Birch £162.75/174/183k and Cedar £180.9k reconcile. Published feasible-minimum scoring, non-delegable sign-off sequence, role capacity and replacement rates are correct.",
            "D": "Central Birch £174k/97 is a specific award recommendation with finance-director signature, 19 April decisions and specialist reservations where needed. Low/growth and confirmed/unconfirmed Cedar yield explicit fallback branches without secretly changing the approved planning basis or method.",
            "E": "Reference quality, release and scale are distinguished; Cedar extra points remain hypothetical until customer confirmation. No demand probabilities, invented risk prices, supplier undertakings or future acceptance are asserted. Direct evidence supports Birch while Alder cleansing remains a Trust burden.",
            "F": "Common cost buildup, annual charging denominator/allowances, exit, weekly allocation, purchases, feasibility gates, scores, assurance conditions and deadline/fallback are linked across all required states.",
            "M": "The central award leads with cost, points and capacity. Tables place the score beside its feasible-price basis; role plans, source references and stated hypothetical conditions make the assumptions easy to inspect.",
            "B": "Coordinator observed growth confirmation changing the winner, a missing Alder purchase excluding its incomplete low-cost plan and reload. Independent interactions cover all 48 path/reference/purchase states and three plan inspectors; displayed winner/cost and reset match. No runtime or internal-anchor errors found."
          },
          "material_failures": [],
          "findings": [],
          "verification_gaps": [
            "This is a separate AI assessment in the same model family, not a human professional check, Claude assessment or evidence of real-world project authority.",
            "Ordinary-browser evidence is the coordinator’s targeted observed journeys; this reviewer ran independent calculations and JSDOM interaction checks, not independent CUA, a full accessibility audit or every device/browser combination."
          ],
          "disposition": "verification only",
          "independent_checks": [
            "48 path/reference/purchase evaluations, normalised costs and scores, plus role capacities/sign-off/test sequencing; 2,528 granular assertions pass.",
            "Baseline Birch£174k/97; growth pendingBirch96.1967 versus Cedar94; confirmedCedar98. Alder growth£212.3k excluded."
          ],
          "artifact_sha256": "7edca985ec1f55a13a9d2b7393e75ff24757bcdf61175a0b3e6ec09eedae6e75",
          "status": "provisional AI assessment",
          "rationale": "AI review disposition: verification only",
          "served_artifact_sha256": "38c1b817a672889c772fc2a28f6cbfcd45cd690bee6d2236c5c8e69f9f141100",
          "review_evidence_id": "essai03-separate-review-01-07"
        },
        "current": {
          "case_id": "BPB2-06",
          "reviewer": "gpt-6-astra · separate assessment agent",
          "method": "Separate source/logic review, independent checks and coordinator ordinary-browser observations; same model family, not a human or Claude assessment; current case analysis is unchanged from the retained first answer",
          "assessed_at": "2026-11-14T18:30:23.049742+00:00",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 4,
            "B": 4
          },
          "reasons": {
            "C": "Independently normalised all three demand paths and both reference states, including all eight specialist-purchase toggles. Alder £148.3/180.3/212.3k, Birch £162.75/174/183k and Cedar £180.9k reconcile. Published feasible-minimum scoring, non-delegable sign-off sequence, role capacity and replacement rates are correct.",
            "D": "Central Birch £174k/97 is a specific award recommendation with finance-director signature, 19 April decisions and specialist reservations where needed. Low/growth and confirmed/unconfirmed Cedar yield explicit fallback branches without secretly changing the approved planning basis or method.",
            "E": "Reference quality, release and scale are distinguished; Cedar extra points remain hypothetical until customer confirmation. No demand probabilities, invented risk prices, supplier undertakings or future acceptance are asserted. Direct evidence supports Birch while Alder cleansing remains a Trust burden.",
            "F": "Common cost buildup, annual charging denominator/allowances, exit, weekly allocation, purchases, feasibility gates, scores, assurance conditions and deadline/fallback are linked across all required states.",
            "M": "The central award leads with cost, points and capacity. Tables place the score beside its feasible-price basis; role plans, source references and stated hypothetical conditions make the assumptions easy to inspect.",
            "B": "Coordinator observed growth confirmation changing the winner, a missing Alder purchase excluding its incomplete low-cost plan and reload. Independent interactions cover all 48 path/reference/purchase states and three plan inspectors; displayed winner/cost and reset match. No runtime or internal-anchor errors found."
          },
          "material_failures": [],
          "findings": [],
          "verification_gaps": [
            "This is a separate AI assessment in the same model family, not a human professional check, Claude assessment or evidence of real-world project authority.",
            "Ordinary-browser evidence is the coordinator’s targeted observed journeys; this reviewer ran independent calculations and JSDOM interaction checks, not independent CUA, a full accessibility audit or every device/browser combination."
          ],
          "disposition": "verification only",
          "independent_checks": [
            "48 path/reference/purchase evaluations, normalised costs and scores, plus role capacities/sign-off/test sequencing; 2,528 granular assertions pass.",
            "Baseline Birch£174k/97; growth pendingBirch96.1967 versus Cedar94; confirmedCedar98. Alder growth£212.3k excluded."
          ],
          "artifact_sha256": "7edca985ec1f55a13a9d2b7393e75ff24757bcdf61175a0b3e6ec09eedae6e75",
          "status": "provisional AI assessment",
          "rationale": "AI review disposition: verification only",
          "served_artifact_sha256": "15bf77c4530e8882502111931b2b472201be9d4c01fb2a675095cff218b35666",
          "review_evidence_id": "essai03-separate-review-01-07"
        }
      },
      "status": "first attempt retained"
    },
    {
      "id": "BPB2-07",
      "self": {
        "initial": {
          "reviewer": "gpt-6-astra · answering agent",
          "method": "Unverified self-assessment before feedback; common assessment/source packaging added afterward",
          "assessed_at": "2026-11-14T18:21:50Z",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 3,
            "B": 3
          },
          "reasons": {
            "C": "All eight offer subsets have complete assignments validated against independent test fixtures for durations, all predecessor links, allowed skills, daily resource exclusivity, closures, rental windows and same-chamber reservation. Hard flight failures remain explicit. Lower bounds establish the necessary bench and why bench plus chamber cannot meet both commitments.",
            "D": "Leads with conditional £12,000 recommendation, £2,000 additional sponsor decision and D3 cutoff; preserves a £5,000 authorised-capacity fallback with D16 customer revision, £7,500 charge and £12,500 consequence. Shows £500 net saving for all three and zero flight margin.",
            "E": "Material facts link to the full supplied records; calculations, assumptions, proposals and unknown approval/customer states are separated. No future consent or actual work completion is invented. Original evidence and initial position remain recoverable.",
            "F": "Includes all twelve packages, dated customer deliveries, people, facilities, dependencies, eight scenario comparisons, useful resource calendar, proof for the principal rejected alternative, customer decision, authority distinctions and a workable fallback.",
            "M": "Decision-first layout, concise financial panels, navigable allocation and source records support the practitioner. Detailed proof and evidence make the page long; calendar and allocation tables require horizontal scrolling on smaller screens.",
            "B": "JSDOM checks pass for initial state, every scenario and dependent result, task inspection, comparison buttons, fallback, filters, reset, evidence anchors, unique IDs and no runtime exceptions. Real browser rendering and keyboard/screen-reader journeys remain unverified."
          },
          "material_failures": [],
          "rationale": "Self-assessment only.",
          "artifact_sha256": "c6e498edeec684c184271d908d2bb277ff32b2ccb632b896fd00f748a2391ea3",
          "verification_gaps": [
            "No real browser rendering, visual inspection or assistive-technology verification has been performed; the coordinator will check the ordinary interface separately.",
            "This is a fixed, validated scenario library and does not optimise arbitrary changes to durations, availability or scope.",
            "Schedules are conditional forecasts; no purchases, funding approvals, customer agreement or real-world certification is asserted."
          ]
        },
        "current": {
          "reviewer": "gpt-6-astra · answering agent",
          "method": "Unverified self-assessment before feedback; common assessment/source packaging added afterward; current case analysis unchanged",
          "assessed_at": "2026-11-14T18:21:50Z",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 3,
            "B": 3
          },
          "reasons": {
            "C": "All eight offer subsets have complete assignments validated against independent test fixtures for durations, all predecessor links, allowed skills, daily resource exclusivity, closures, rental windows and same-chamber reservation. Hard flight failures remain explicit. Lower bounds establish the necessary bench and why bench plus chamber cannot meet both commitments.",
            "D": "Leads with conditional £12,000 recommendation, £2,000 additional sponsor decision and D3 cutoff; preserves a £5,000 authorised-capacity fallback with D16 customer revision, £7,500 charge and £12,500 consequence. Shows £500 net saving for all three and zero flight margin.",
            "E": "Material facts link to the full supplied records; calculations, assumptions, proposals and unknown approval/customer states are separated. No future consent or actual work completion is invented. Original evidence and initial position remain recoverable.",
            "F": "Includes all twelve packages, dated customer deliveries, people, facilities, dependencies, eight scenario comparisons, useful resource calendar, proof for the principal rejected alternative, customer decision, authority distinctions and a workable fallback.",
            "M": "Decision-first layout, concise financial panels, navigable allocation and source records support the practitioner. Detailed proof and evidence make the page long; calendar and allocation tables require horizontal scrolling on smaller screens.",
            "B": "JSDOM checks pass for initial state, every scenario and dependent result, task inspection, comparison buttons, fallback, filters, reset, evidence anchors, unique IDs and no runtime exceptions. Real browser rendering and keyboard/screen-reader journeys remain unverified."
          },
          "material_failures": [],
          "rationale": "Self-assessment only.",
          "artifact_sha256": "c6e498edeec684c184271d908d2bb277ff32b2ccb632b896fd00f748a2391ea3",
          "verification_gaps": [
            "No real browser rendering, visual inspection or assistive-technology verification has been performed; the coordinator will check the ordinary interface separately.",
            "This is a fixed, validated scenario library and does not optimise arbitrary changes to durations, availability or scope.",
            "Schedules are conditional forecasts; no purchases, funding approvals, customer agreement or real-world certification is asserted."
          ]
        }
      },
      "checked": {
        "initial": {
          "case_id": "BPB2-07",
          "reviewer": "gpt-6-astra · separate assessment agent",
          "method": "Separate source/logic review, independent checks and coordinator ordinary-browser observations; same model family, not a human or Claude assessment",
          "assessed_at": "2026-11-14T18:30:23.049742+00:00",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 4,
            "B": 4
          },
          "reasons": {
            "C": "Independently audited all eight twelve-task witnesses for duration, precedence, skills, staff/facility overlap, closures, rental windows and same chamber. The retained B fallback G13/K16 is feasible; BTF G13/K12 and £12k total are correct. B/BT/BF K16 costs and no-bench G14 failures reconcile. An independently validated G12/K16 B witness provides an optional extra margin day.",
            "D": "The page protects the hard flight commitment, recommends BTF with a specific £2k funding decision by D3 and keeps the £5k B fallback executable. The £12k versus £12.5k purchase-plus-lateness comparison and outstanding customer agreement are clear. The free fallback-margin improvement is recorded as minor; the task permits alternative feasible schedules and does not require its exact key witness.",
            "E": "Package-lead claims are tested against named resource records. Approval, reservation, customer acceptance and successful execution stay unknown; no unpriced flight penalty or risk probability is invented. The page correctly labels its schedules as deterministic self-checked plans, not certification.",
            "F": "All twelve assignments, dependencies, dates, daily lanes, task inspections, required purchase comparisons, lower-bound reasoning, authority gaps, charge forecasts and fallback deadline form a coherent recovery brief.",
            "M": "The extra £2k decision and fallback lead. A compact scenario comparison, task inspector and resource calendar provide detail. Zero submitted flight margin, one-day preferred endurance margin, unagreed fallback date and distinct purchase authority are plainly stated.",
            "B": "Coordinator observed B/BT/none switches, dates/costs, unused mobile chamber, facility filter and reset/reload. Independent interaction checks cover eight selections, twelve task selections for each, fallback and reset; all schedule functions and anchors run without error. Ordinary observations support the main required journeys, not universal browser/device certification."
          },
          "material_failures": [],
          "findings": [
            {
              "severity": "minor",
              "description": "Bench-only fallback can gain one day of flight margin at no extra cost while retaining endurance D16. The retained G13/K16 schedule remains feasible and correctly reports its own zero margin.",
              "evidence": "case07-alternate-result.json validates C4–5 Ben/owned bench; D6–7 Cara/rental; E8–9 Ada/owned; F10–11 Cara/owned; G12; H8–10 Dev/rental; L4–5 Ada; I11–12 Ben/owned chamber; J13–15 Cara; K16.",
              "proposed_repair": "Optional bounded refinement: offer this G12/K16 allocation as the B fallback and update its displayed flight margin/date. Consider the same free schedule for BT/BF inspections. Retain the original first artifact; no repair has been performed."
            }
          ],
          "verification_gaps": [
            "This is a separate AI assessment in the same model family, not a human professional check, Claude assessment or evidence of real-world project authority.",
            "Ordinary-browser evidence is the coordinator’s targeted observed journeys; this reviewer ran independent calculations and JSDOM interaction checks, not independent CUA, a full accessibility audit or every device/browser combination."
          ],
          "disposition": "verification only",
          "independent_checks": [
            "Eight full schedule/resource/cost witnesses checked; 1,224 granular assertions pass. BTFG13/K12£12k; BG13/K16£12.5k; BT£16.5k without useful mobile use.",
            "After candidate-key consultation independently validated an alternate B-only G12/K16 witness in case07-alternate-check.py:108 checks pass; same£12.5k consequence and one flight-margin day."
          ],
          "artifact_sha256": "c6e498edeec684c184271d908d2bb277ff32b2ccb632b896fd00f748a2391ea3",
          "status": "provisional AI assessment",
          "rationale": "AI review disposition: verification only",
          "served_artifact_sha256": "9f894ff3be9fca185037319d160196d03cb7366a73b46653702b7b563b5f14de",
          "review_evidence_id": "essai03-separate-review-01-07"
        },
        "current": {
          "case_id": "BPB2-07",
          "reviewer": "gpt-6-astra · separate assessment agent",
          "method": "Separate source/logic review, independent checks and coordinator ordinary-browser observations; same model family, not a human or Claude assessment; current case analysis is unchanged from the retained first answer",
          "assessed_at": "2026-11-14T18:30:23.049742+00:00",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 4,
            "B": 4
          },
          "reasons": {
            "C": "Independently audited all eight twelve-task witnesses for duration, precedence, skills, staff/facility overlap, closures, rental windows and same chamber. The retained B fallback G13/K16 is feasible; BTF G13/K12 and £12k total are correct. B/BT/BF K16 costs and no-bench G14 failures reconcile. An independently validated G12/K16 B witness provides an optional extra margin day.",
            "D": "The page protects the hard flight commitment, recommends BTF with a specific £2k funding decision by D3 and keeps the £5k B fallback executable. The £12k versus £12.5k purchase-plus-lateness comparison and outstanding customer agreement are clear. The free fallback-margin improvement is recorded as minor; the task permits alternative feasible schedules and does not require its exact key witness.",
            "E": "Package-lead claims are tested against named resource records. Approval, reservation, customer acceptance and successful execution stay unknown; no unpriced flight penalty or risk probability is invented. The page correctly labels its schedules as deterministic self-checked plans, not certification.",
            "F": "All twelve assignments, dependencies, dates, daily lanes, task inspections, required purchase comparisons, lower-bound reasoning, authority gaps, charge forecasts and fallback deadline form a coherent recovery brief.",
            "M": "The extra £2k decision and fallback lead. A compact scenario comparison, task inspector and resource calendar provide detail. Zero submitted flight margin, one-day preferred endurance margin, unagreed fallback date and distinct purchase authority are plainly stated.",
            "B": "Coordinator observed B/BT/none switches, dates/costs, unused mobile chamber, facility filter and reset/reload. Independent interaction checks cover eight selections, twelve task selections for each, fallback and reset; all schedule functions and anchors run without error. Ordinary observations support the main required journeys, not universal browser/device certification."
          },
          "material_failures": [],
          "findings": [
            {
              "severity": "minor",
              "description": "Bench-only fallback can gain one day of flight margin at no extra cost while retaining endurance D16. The retained G13/K16 schedule remains feasible and correctly reports its own zero margin.",
              "evidence": "case07-alternate-result.json validates C4–5 Ben/owned bench; D6–7 Cara/rental; E8–9 Ada/owned; F10–11 Cara/owned; G12; H8–10 Dev/rental; L4–5 Ada; I11–12 Ben/owned chamber; J13–15 Cara; K16.",
              "proposed_repair": "Optional bounded refinement: offer this G12/K16 allocation as the B fallback and update its displayed flight margin/date. Consider the same free schedule for BT/BF inspections. Retain the original first artifact; no repair has been performed."
            }
          ],
          "verification_gaps": [
            "This is a separate AI assessment in the same model family, not a human professional check, Claude assessment or evidence of real-world project authority.",
            "Ordinary-browser evidence is the coordinator’s targeted observed journeys; this reviewer ran independent calculations and JSDOM interaction checks, not independent CUA, a full accessibility audit or every device/browser combination."
          ],
          "disposition": "verification only",
          "independent_checks": [
            "Eight full schedule/resource/cost witnesses checked; 1,224 granular assertions pass. BTFG13/K12£12k; BG13/K16£12.5k; BT£16.5k without useful mobile use.",
            "After candidate-key consultation independently validated an alternate B-only G12/K16 witness in case07-alternate-check.py:108 checks pass; same£12.5k consequence and one flight-margin day."
          ],
          "artifact_sha256": "c6e498edeec684c184271d908d2bb277ff32b2ccb632b896fd00f748a2391ea3",
          "status": "provisional AI assessment",
          "rationale": "AI review disposition: verification only",
          "served_artifact_sha256": "dfaef025d1a7a21d17c2632fd6146cc5c4e9bbe0e58d57ed3da61234ebb3c546",
          "review_evidence_id": "essai03-separate-review-01-07"
        }
      },
      "status": "first attempt retained"
    },
    {
      "id": "BPB2-08",
      "self": {
        "initial": {
          "reviewer": "gpt-6-astra · answering agent",
          "method": "Unverified self-assessment before feedback; common assessment/source packaging added afterward",
          "assessed_at": "2026-11-14T17:44:45Z",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 3,
            "B": 2
          },
          "reasons": {
            "C": "Reconciles the £59,500 additional claim to £35,400; retains CE04 once, corrects full omission and fee base, models the D9 unchanged-design counterfactual and concurrency exactly, and keeps the £1,500 existing entitlement separate. All three calculated option totals and dates passed explicit DOM assertions.",
            "D": "Leads with a conditional normal-acoustic recommendation and a timed original-design fallback, identifies required authority and tenant agreement, and quantifies freight's £6,900 extra contract cost, £4,000 accommodation saving and £2,900 net premium without inventing acoustic value.",
            "E": "All nine supplied records are recoverable with local source links; facts, paid authority, contingent forecasts and proposed instructions are distinguished. No tenant funding, variation acceptance or sponsor signature is assumed.",
            "F": "Includes three dated critical sequences, delay attribution, detailed quotation reconciliation, selected-option whole event/CE04/additional/existing entitlement/forecast/tenant results, expiry and fallback, reset and reload behaviour.",
            "M": "Practical memo structure, clear headline and explicit recommended next instruction; the detailed reconciliation and source archive make the page long. Rendered layout has not been inspected.",
            "B": "Eight JSDOM functional checks passed, including option controls, reset, fresh load, evidence expansion and no script errors. Real-browser launch was unavailable, so rendering, accessibility and responsive behaviour remain unverified."
          },
          "material_failures": [],
          "rationale": "PM dimensions meet the threshold, but B remains below 3 pending ordinary browser verification. This is a self-assessment, not independent validation.",
          "artifact_sha256": "8b55c18a3f057b6fcdf3ad3b160622d4899f134d8ff28fb124e3a354cecd61b1",
          "verification_gaps": []
        },
        "current": {
          "reviewer": "gpt-6-astra · answering agent",
          "method": "Unverified self-assessment before feedback; common assessment/source packaging added afterward; current case analysis unchanged",
          "assessed_at": "2026-11-14T17:44:45Z",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 3,
            "B": 2
          },
          "reasons": {
            "C": "Reconciles the £59,500 additional claim to £35,400; retains CE04 once, corrects full omission and fee base, models the D9 unchanged-design counterfactual and concurrency exactly, and keeps the £1,500 existing entitlement separate. All three calculated option totals and dates passed explicit DOM assertions.",
            "D": "Leads with a conditional normal-acoustic recommendation and a timed original-design fallback, identifies required authority and tenant agreement, and quantifies freight's £6,900 extra contract cost, £4,000 accommodation saving and £2,900 net premium without inventing acoustic value.",
            "E": "All nine supplied records are recoverable with local source links; facts, paid authority, contingent forecasts and proposed instructions are distinguished. No tenant funding, variation acceptance or sponsor signature is assumed.",
            "F": "Includes three dated critical sequences, delay attribution, detailed quotation reconciliation, selected-option whole event/CE04/additional/existing entitlement/forecast/tenant results, expiry and fallback, reset and reload behaviour.",
            "M": "Practical memo structure, clear headline and explicit recommended next instruction; the detailed reconciliation and source archive make the page long. Rendered layout has not been inspected.",
            "B": "Eight JSDOM functional checks passed, including option controls, reset, fresh load, evidence expansion and no script errors. Real-browser launch was unavailable, so rendering, accessibility and responsive behaviour remain unverified."
          },
          "material_failures": [],
          "rationale": "PM dimensions meet the threshold, but B remains below 3 pending ordinary browser verification. This is a self-assessment, not independent validation.",
          "artifact_sha256": "8b55c18a3f057b6fcdf3ad3b160622d4899f134d8ff28fb124e3a354cecd61b1",
          "verification_gaps": []
        }
      },
      "checked": {
        "initial": {
          "reviewer": "gpt-6-astra · separate assessment agent",
          "method": "Separate source/logic review, independent checks and coordinator ordinary-browser observations; same model family, not a human or Claude assessment",
          "assessed_at": "2026-11-14T18:49:43.095990Z",
          "case_id": "BPB2-08",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 4,
            "B": 4
          },
          "reasons": {
            "C": "Independently recomputed all three costs and programmes: normal D20/27 August, £42,000 whole event, £35,400 remaining, £463,500 forecast; expedited D18, £42,300 remaining, £470,400 forecast; original D18, £2,200 remaining, £430,300 forecast. CE04 is deducted once and the separate £1,500 existing entitlement and £10,000/£6,000 tenant costs reconcile.",
            "D": "Prioritises normal acoustics only with tenant price/delay/funding acceptance, prices expedited two-day benefit at £2,900 combined incremental cost, and gives an explicit original-design fallback before 17:00 with Mara/Helena authority separated.",
            "E": "Preserves performed/paid CE04, live original order, deposit/refund position, verified concurrency and unresolved tenant/sponsor acceptance. Source IDs and full original records accompany the material claims; no invented benefit valuation or law.",
            "F": "Reconciliation, paid-work deduction, whole event versus remaining authority, three sequences, unchanged-design counterfactual, concurrency and compensation, separate tenant economics and deadline/approval route form a coherent decision.",
            "M": "Decision and 17:00 expiry lead the page; option table and concise trade-off expose what freight buys. Dates, working-day rules, current authority versus forecast, and tenant versus contract cost prevent likely misreadings.",
            "B": "Coordinator ordinary-browser changes to expedited/original and reload worked. Independent DOM tests confirmed all three selected dependent monetary values, unchanged £426,600/£1,500 lines, source-link targets and reset. The finite option control needs no free-input rejection."
          },
          "material_failures": [],
          "findings": [],
          "verification_gaps": [
            "This is a separate AI assessment within the same model family, not Claude or human professional validation.",
            "Native browser evidence is the coordinator’s bounded ordinary-browser observations; the assessor independently ran JSDOM checks, which do not establish comprehensive rendering, keyboard or assistive-technology coverage."
          ],
          "disposition": "verification only",
          "independent_checks": [
            "Rebuilt start=max(6,9,8,delivery), ten-day sequence and D18/D20 completion; pre-existing three extension days, one payable day; normal two incremental days.",
            "Recomputed gross events £42,000/£48,900/£8,800, additional instructions £35,400/£42,300/£2,200 and combined economics £473,500/£476,400/£436,300.",
            "All three option calculations and displayed cost propagation agreed in independent DOM check; coordinator browser observations support route/reload.",
            "Reviewer DOM outcomes: 6 passed checks, 0 failed assertions. See saved checking code and results; these are not the author’s self-tests."
          ],
          "artifact_sha256": "8b55c18a3f057b6fcdf3ad3b160622d4899f134d8ff28fb124e3a354cecd61b1",
          "status": "provisional AI assessment",
          "rationale": "AI review disposition: verification only",
          "served_artifact_sha256": "13d6883d7eaaf19b7ff4df8ed8b9ca3e54b8d8e7b8cddf96ecc40b78c930e1cd",
          "review_evidence_id": "essai03-separate-review-08-14"
        },
        "current": {
          "reviewer": "gpt-6-astra · separate assessment agent",
          "method": "Separate source/logic review, independent checks and coordinator ordinary-browser observations; same model family, not a human or Claude assessment; current case analysis is unchanged from the retained first answer",
          "assessed_at": "2026-11-14T18:49:43.095990Z",
          "case_id": "BPB2-08",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 4,
            "B": 4
          },
          "reasons": {
            "C": "Independently recomputed all three costs and programmes: normal D20/27 August, £42,000 whole event, £35,400 remaining, £463,500 forecast; expedited D18, £42,300 remaining, £470,400 forecast; original D18, £2,200 remaining, £430,300 forecast. CE04 is deducted once and the separate £1,500 existing entitlement and £10,000/£6,000 tenant costs reconcile.",
            "D": "Prioritises normal acoustics only with tenant price/delay/funding acceptance, prices expedited two-day benefit at £2,900 combined incremental cost, and gives an explicit original-design fallback before 17:00 with Mara/Helena authority separated.",
            "E": "Preserves performed/paid CE04, live original order, deposit/refund position, verified concurrency and unresolved tenant/sponsor acceptance. Source IDs and full original records accompany the material claims; no invented benefit valuation or law.",
            "F": "Reconciliation, paid-work deduction, whole event versus remaining authority, three sequences, unchanged-design counterfactual, concurrency and compensation, separate tenant economics and deadline/approval route form a coherent decision.",
            "M": "Decision and 17:00 expiry lead the page; option table and concise trade-off expose what freight buys. Dates, working-day rules, current authority versus forecast, and tenant versus contract cost prevent likely misreadings.",
            "B": "Coordinator ordinary-browser changes to expedited/original and reload worked. Independent DOM tests confirmed all three selected dependent monetary values, unchanged £426,600/£1,500 lines, source-link targets and reset. The finite option control needs no free-input rejection."
          },
          "material_failures": [],
          "findings": [],
          "verification_gaps": [
            "This is a separate AI assessment within the same model family, not Claude or human professional validation.",
            "Native browser evidence is the coordinator’s bounded ordinary-browser observations; the assessor independently ran JSDOM checks, which do not establish comprehensive rendering, keyboard or assistive-technology coverage."
          ],
          "disposition": "verification only",
          "independent_checks": [
            "Rebuilt start=max(6,9,8,delivery), ten-day sequence and D18/D20 completion; pre-existing three extension days, one payable day; normal two incremental days.",
            "Recomputed gross events £42,000/£48,900/£8,800, additional instructions £35,400/£42,300/£2,200 and combined economics £473,500/£476,400/£436,300.",
            "All three option calculations and displayed cost propagation agreed in independent DOM check; coordinator browser observations support route/reload.",
            "Reviewer DOM outcomes: 6 passed checks, 0 failed assertions. See saved checking code and results; these are not the author’s self-tests."
          ],
          "artifact_sha256": "8b55c18a3f057b6fcdf3ad3b160622d4899f134d8ff28fb124e3a354cecd61b1",
          "status": "provisional AI assessment",
          "rationale": "AI review disposition: verification only",
          "served_artifact_sha256": "42531cfca0122c6dbe9fd62b1d7b7484a8d713fa4c9de9aa50b0843a8eade91c",
          "review_evidence_id": "essai03-separate-review-08-14"
        }
      },
      "status": "first attempt retained"
    },
    {
      "id": "BPB2-09",
      "self": {
        "initial": {
          "reviewer": "gpt-6-astra · answering agent",
          "method": "Unverified self-assessment before feedback; common assessment/source packaging added afterward",
          "assessed_at": "2026-11-14T18:27:10Z",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 3,
            "B": 3
          },
          "reasons": {
            "C": "Whole-package selection, H once, hard role/week dependencies and A/S interaction are explicit. All 128 optional-scope/specialist combinations match a separate direct feasibility predicate; allocations preserve effort and allowed windows.",
            "D": "Recommends S+X with specialist and compares three funding/capacity-feasible choices, with costs, dates, staffing, migration limits and acceptance consequences. Failed selections show concrete blockers.",
            "E": "Original 09-R1–R8 records are retained and linked. Observed baseline, conditional projections, forecasts and proposed owners are separated; no signatures, completed tests, appointments or full migration are claimed.",
            "F": "Interactive scope, budget, per-role/week allocation, dependencies, record partitions, operating owners, bridge expiry, preceding decision and recovery/cutover gates are covered. Reset and reload preserve the original dated position.",
            "M": "Practical recommendation leads; comparison, navigable sections and source disclosures aid review. The complete page is long and dense, especially the acceptance and evidence sections.",
            "B": "Eleven JSDOM checks passed, including all control journeys and 128 combinations, without script errors. Real-browser visual, keyboard and screen-reader inspection was not available and remains unverified."
          },
          "material_failures": [],
          "rationale": "Self-assessment only.",
          "artifact_sha256": "70cb09db1d4169eb5e6cac59cd73dead234eb5a8019fe68ba121583587ef2181",
          "verification_gaps": [
            "No real-browser rendering or ordinary keyboard/screen-reader walkthrough performed.",
            "These synthetic planning results do not constitute stakeholder approval, owner appointment, procurement or completed acceptance tests."
          ]
        },
        "current": {
          "reviewer": "gpt-6-astra · answering agent",
          "method": "Unverified self-assessment before feedback; common assessment/source packaging added afterward; current case analysis unchanged",
          "assessed_at": "2026-11-14T18:27:10Z",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 3,
            "B": 3
          },
          "reasons": {
            "C": "Whole-package selection, H once, hard role/week dependencies and A/S interaction are explicit. All 128 optional-scope/specialist combinations match a separate direct feasibility predicate; allocations preserve effort and allowed windows.",
            "D": "Recommends S+X with specialist and compares three funding/capacity-feasible choices, with costs, dates, staffing, migration limits and acceptance consequences. Failed selections show concrete blockers.",
            "E": "Original 09-R1–R8 records are retained and linked. Observed baseline, conditional projections, forecasts and proposed owners are separated; no signatures, completed tests, appointments or full migration are claimed.",
            "F": "Interactive scope, budget, per-role/week allocation, dependencies, record partitions, operating owners, bridge expiry, preceding decision and recovery/cutover gates are covered. Reset and reload preserve the original dated position.",
            "M": "Practical recommendation leads; comparison, navigable sections and source disclosures aid review. The complete page is long and dense, especially the acceptance and evidence sections.",
            "B": "Eleven JSDOM checks passed, including all control journeys and 128 combinations, without script errors. Real-browser visual, keyboard and screen-reader inspection was not available and remains unverified."
          },
          "material_failures": [],
          "rationale": "Self-assessment only.",
          "artifact_sha256": "70cb09db1d4169eb5e6cac59cd73dead234eb5a8019fe68ba121583587ef2181",
          "verification_gaps": [
            "No real-browser rendering or ordinary keyboard/screen-reader walkthrough performed.",
            "These synthetic planning results do not constitute stakeholder approval, owner appointment, procurement or completed acceptance tests."
          ]
        }
      },
      "checked": {
        "initial": {
          "reviewer": "gpt-6-astra · separate assessment agent",
          "method": "Separate source/logic review, independent checks and coordinator ordinary-browser observations; same model family, not a human or Claude assessment",
          "assessed_at": "2026-11-14T18:49:43.095990Z",
          "case_id": "BPB2-09",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 4,
            "B": 4
          },
          "reasons": {
            "C": "Recomputed every one of 128 scope/specialist combinations using independent week-subset capacity inequalities and checked the submitted allocations conserve effort within allowed weeks. All feasibility flags, cash and totals agreed. The three supplied routes fit; A+S W5 development, X+Y W3–4 data, R/K and reserve violations are exposed.",
            "D": "Recommends service plus full mandate repair with a £1,200 specialist; compares finance/acknowledgement and consent routes with explicit capacity, cost, record and acceptance consequences. Gives a funded decision process and warns that a failed X does not produce a free fallback.",
            "E": "Records and prospective post-test counts are separate; quarantine is not consent, finance bridge acceptance in principle is not release approval, draft owner nomination is not appointment, and read-only legacy availability is preserved. Deferred report plan is honestly unagreed.",
            "F": "Covers weekly roles, shared H once, A/S interaction, migrations, mandate money/states, bridge funding and daily ownership, W4 interface versus W5 finance sign-off, rollback baseline and 3 December proposed decision ahead of 10 December expiry.",
            "M": "The recommended release and trade-off lead the page, and a compact three-choice comparison precedes controls and detailed allocation. Scope consequences and conditional status are close to claims; new-record accounting cannot readily be mistaken for complete consent cleanup.",
            "B": "Coordinator ordinary-browser additions/removals exposed W5 and W4 overloads and reset cleared nomination. Independent tests checked all 128 masks, three preset routes, source anchors and reset. Required selectable scope states propagate into plans and records; infeasible proposals are visibly invalid."
          },
          "material_failures": [],
          "findings": [
            {
              "severity": "observation",
              "description": "The answer deliberately omits optional L from its three choices, unlike some candidate-key witnesses. This is permitted: free read-only legacy remains available, and the answer explicitly defers L.",
              "evidence": "Source pack, answer and candidate key compared after independent interpretation; see independent-interpretation.md and independent checks.",
              "proposed_repair": "No repair required on this point; retain the defensible alternative rather than force key wording or allocation."
            }
          ],
          "verification_gaps": [
            "This is a separate AI assessment within the same model family, not Claude or human professional validation.",
            "Native browser evidence is the coordinator’s bounded ordinary-browser observations; the assessor independently ran JSDOM checks, which do not establish comprehensive rendering, keyboard or assistive-technology coverage."
          ],
          "disposition": "verification only",
          "independent_checks": [
            "128 combinations independently classified using all subsets of five planning periods for each role; 16 fit both capacity and £1,800 reserve, matching answer.",
            "All allocation rows conserve package effort and use permitted role/week windows. Three offered routes cost £1,200/£800/£1,700.",
            "Independently reconciled 10,000 contacts, 640 mandates and £71,250; Y changes to 7,600/1,500/900; X changes active-new 584 to620.",
            "Reviewer DOM outcomes: 134 passed checks, 0 failed assertions. See saved checking code and results; these are not the author’s self-tests."
          ],
          "artifact_sha256": "70cb09db1d4169eb5e6cac59cd73dead234eb5a8019fe68ba121583587ef2181",
          "status": "provisional AI assessment",
          "rationale": "AI review disposition: verification only",
          "served_artifact_sha256": "249dfafa0489e9cd463e3bacfeffafdf24fd9d46391049cd097231ec78725f69",
          "review_evidence_id": "essai03-separate-review-08-14"
        },
        "current": {
          "reviewer": "gpt-6-astra · separate assessment agent",
          "method": "Separate source/logic review, independent checks and coordinator ordinary-browser observations; same model family, not a human or Claude assessment; current case analysis is unchanged from the retained first answer",
          "assessed_at": "2026-11-14T18:49:43.095990Z",
          "case_id": "BPB2-09",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 4,
            "B": 4
          },
          "reasons": {
            "C": "Recomputed every one of 128 scope/specialist combinations using independent week-subset capacity inequalities and checked the submitted allocations conserve effort within allowed weeks. All feasibility flags, cash and totals agreed. The three supplied routes fit; A+S W5 development, X+Y W3–4 data, R/K and reserve violations are exposed.",
            "D": "Recommends service plus full mandate repair with a £1,200 specialist; compares finance/acknowledgement and consent routes with explicit capacity, cost, record and acceptance consequences. Gives a funded decision process and warns that a failed X does not produce a free fallback.",
            "E": "Records and prospective post-test counts are separate; quarantine is not consent, finance bridge acceptance in principle is not release approval, draft owner nomination is not appointment, and read-only legacy availability is preserved. Deferred report plan is honestly unagreed.",
            "F": "Covers weekly roles, shared H once, A/S interaction, migrations, mandate money/states, bridge funding and daily ownership, W4 interface versus W5 finance sign-off, rollback baseline and 3 December proposed decision ahead of 10 December expiry.",
            "M": "The recommended release and trade-off lead the page, and a compact three-choice comparison precedes controls and detailed allocation. Scope consequences and conditional status are close to claims; new-record accounting cannot readily be mistaken for complete consent cleanup.",
            "B": "Coordinator ordinary-browser additions/removals exposed W5 and W4 overloads and reset cleared nomination. Independent tests checked all 128 masks, three preset routes, source anchors and reset. Required selectable scope states propagate into plans and records; infeasible proposals are visibly invalid."
          },
          "material_failures": [],
          "findings": [
            {
              "severity": "observation",
              "description": "The answer deliberately omits optional L from its three choices, unlike some candidate-key witnesses. This is permitted: free read-only legacy remains available, and the answer explicitly defers L.",
              "evidence": "Source pack, answer and candidate key compared after independent interpretation; see independent-interpretation.md and independent checks.",
              "proposed_repair": "No repair required on this point; retain the defensible alternative rather than force key wording or allocation."
            }
          ],
          "verification_gaps": [
            "This is a separate AI assessment within the same model family, not Claude or human professional validation.",
            "Native browser evidence is the coordinator’s bounded ordinary-browser observations; the assessor independently ran JSDOM checks, which do not establish comprehensive rendering, keyboard or assistive-technology coverage."
          ],
          "disposition": "verification only",
          "independent_checks": [
            "128 combinations independently classified using all subsets of five planning periods for each role; 16 fit both capacity and £1,800 reserve, matching answer.",
            "All allocation rows conserve package effort and use permitted role/week windows. Three offered routes cost £1,200/£800/£1,700.",
            "Independently reconciled 10,000 contacts, 640 mandates and £71,250; Y changes to 7,600/1,500/900; X changes active-new 584 to620.",
            "Reviewer DOM outcomes: 134 passed checks, 0 failed assertions. See saved checking code and results; these are not the author’s self-tests."
          ],
          "artifact_sha256": "70cb09db1d4169eb5e6cac59cd73dead234eb5a8019fe68ba121583587ef2181",
          "status": "provisional AI assessment",
          "rationale": "AI review disposition: verification only",
          "served_artifact_sha256": "fdbb99223b5e55fd52a65df956eff0f43a323946d38ae365576e8b7dc231bb53",
          "review_evidence_id": "essai03-separate-review-08-14"
        }
      },
      "status": "first attempt retained"
    },
    {
      "id": "BPB2-10",
      "self": {
        "initial": {
          "reviewer": "gpt-6-astra · answering agent",
          "method": "Unverified self-assessment before feedback; common assessment/source packaging added afterward",
          "assessed_at": "2026-11-14T18:26:52Z",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 3,
            "B": 3
          },
          "reasons": {
            "C": "Reconstructs twelve authorised units after split/merge. Both numerators are eight, with W102 replaced by W110 only in the restatement. Approved event-date aliases and position revisions are applied; failed, wrong-position, unknown-identity and post-cutoff work are excluded.",
            "D": "Leads with a practical PM publication recommendation, shows explicit accepted lists and membership delta, and provides a 12 May queue with accountable authority roles and closure evidence.",
            "E": "Separates issue time, event time, original cutoff evidence, verified restatement and subsequent events. Full supplied sources remain readable. No invented U-99 match, coordinate approval, defect destination or future outcome.",
            "F": "Includes all twelve current units, claim normalisation, split/merge/alias/correction traces, unresolved filters, later-work note, assertion assessment, action queue, original sources, reset and gallery link.",
            "M": "Professional reporting-record layout and source-linked plain-language reasons are implemented. Some tables are necessarily dense; actual rendered appearance and mobile ergonomics have not been independently viewed.",
            "B": "Twenty-five JSDOM interaction and structure checks pass with no script errors. All required controls change dependent content. Real browser rendering, keyboard focus visibility and physical navigation were not tested here."
          },
          "material_failures": [],
          "rationale": "Self-assessment only.",
          "artifact_sha256": "72f2dedaa839cc1276a998b3676a9982c2fc79babe661a831e0e07a0fa0b6b7d",
          "verification_gaps": []
        },
        "current": {
          "reviewer": "gpt-6-astra · answering agent",
          "method": "Unverified self-assessment before feedback; common assessment/source packaging added afterward; current case analysis unchanged",
          "assessed_at": "2026-11-14T18:26:52Z",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 3,
            "B": 3
          },
          "reasons": {
            "C": "Reconstructs twelve authorised units after split/merge. Both numerators are eight, with W102 replaced by W110 only in the restatement. Approved event-date aliases and position revisions are applied; failed, wrong-position, unknown-identity and post-cutoff work are excluded.",
            "D": "Leads with a practical PM publication recommendation, shows explicit accepted lists and membership delta, and provides a 12 May queue with accountable authority roles and closure evidence.",
            "E": "Separates issue time, event time, original cutoff evidence, verified restatement and subsequent events. Full supplied sources remain readable. No invented U-99 match, coordinate approval, defect destination or future outcome.",
            "F": "Includes all twelve current units, claim normalisation, split/merge/alias/correction traces, unresolved filters, later-work note, assertion assessment, action queue, original sources, reset and gallery link.",
            "M": "Professional reporting-record layout and source-linked plain-language reasons are implemented. Some tables are necessarily dense; actual rendered appearance and mobile ergonomics have not been independently viewed.",
            "B": "Twenty-five JSDOM interaction and structure checks pass with no script errors. All required controls change dependent content. Real browser rendering, keyboard focus visibility and physical navigation were not tested here."
          },
          "material_failures": [],
          "rationale": "Self-assessment only.",
          "artifact_sha256": "72f2dedaa839cc1276a998b3676a9982c2fc79babe661a831e0e07a0fa0b6b7d",
          "verification_gaps": []
        }
      },
      "checked": {
        "initial": {
          "reviewer": "gpt-6-astra · separate assessment agent",
          "method": "Separate source/logic review, independent checks and coordinator ordinary-browser observations; same model family, not a human or Claude assessment",
          "assessed_at": "2026-11-14T18:49:43.095990Z",
          "case_id": "BPB2-10",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 4,
            "B": 4
          },
          "reasons": {
            "C": "Reconstructed 12 authorised units, 11 claim-covered units and both eight-unit acceptance sets. W102 is removed and W110 added only in the restatement; W112 remains later work. Approved W106 V2 and event-specific OLD-17 mapping are applied; W107 is 30m out and U-99 stays unresolved.",
            "D": "Recommends publishing both historical views with an explicit membership-change statement and a specific latest-evidence action queue for W102, W104B, W107, U-99 and publication. Owner, inspection lead, contractor and PM responsibilities respect their powers.",
            "E": "Issue/effective/event distinctions remain visible; original plausible I2 is not retrospectively treated as knowably wrong. Parent/component/history and unknown identity remain inspectable without fabricated matches or erasure of superseded evidence.",
            "F": "Current register, split/merge/alias/correction traces, exceptions, claim-count reconciliation, competing assertions and subsequent-events note coherently fulfil both historical views and follow-through.",
            "M": "Same percentage/different membership is the prominent argument. Side-by-side sets and changed-unit filter prevent an unchanged66.7% headline from concealing the correction; source detail and action queue are easy to reach.",
            "B": "Coordinator browser showed the unchanged8/12 with W102/W110 swap, U-99 zero-match and OLD-17 trace. Independent DOM check verified both sets, all six traces, changed filter, source targets and reset. No numerical free input or unsupported identity edit is exposed."
          },
          "material_failures": [],
          "findings": [],
          "verification_gaps": [
            "This is a separate AI assessment within the same model family, not Claude or human professional validation.",
            "Native browser evidence is the coordinator’s bounded ordinary-browser observations; the assessor independently ran JSDOM checks, which do not establish comprehensive rendering, keyboard or assistive-technology coverage."
          ],
          "disposition": "verification only",
          "independent_checks": [
            "Authorised denominator12−1+2−2+1=12; current claim coverage16−1duplicate−1parent−1unresolved−2merge consolidation=11.",
            "As-known accepted set W101,W102,W103,W104A,W105,W106,W1089,W111; restated removesW102/addsW110. DOM sets agree.",
            "W102 distance5m; W107 distance30m; later W112 event excluded. Six lineage disclosures and unresolved search tested.",
            "Reviewer DOM outcomes: 13 passed checks, 0 failed assertions. See saved checking code and results; these are not the author’s self-tests."
          ],
          "artifact_sha256": "72f2dedaa839cc1276a998b3676a9982c2fc79babe661a831e0e07a0fa0b6b7d",
          "status": "provisional AI assessment",
          "rationale": "AI review disposition: verification only",
          "served_artifact_sha256": "4e201f099397e5036433a8e05b5f23074f1cce6dec0c81bad277c1eb01fd6643",
          "review_evidence_id": "essai03-separate-review-08-14"
        },
        "current": {
          "reviewer": "gpt-6-astra · separate assessment agent",
          "method": "Separate source/logic review, independent checks and coordinator ordinary-browser observations; same model family, not a human or Claude assessment; current case analysis is unchanged from the retained first answer",
          "assessed_at": "2026-11-14T18:49:43.095990Z",
          "case_id": "BPB2-10",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 4,
            "B": 4
          },
          "reasons": {
            "C": "Reconstructed 12 authorised units, 11 claim-covered units and both eight-unit acceptance sets. W102 is removed and W110 added only in the restatement; W112 remains later work. Approved W106 V2 and event-specific OLD-17 mapping are applied; W107 is 30m out and U-99 stays unresolved.",
            "D": "Recommends publishing both historical views with an explicit membership-change statement and a specific latest-evidence action queue for W102, W104B, W107, U-99 and publication. Owner, inspection lead, contractor and PM responsibilities respect their powers.",
            "E": "Issue/effective/event distinctions remain visible; original plausible I2 is not retrospectively treated as knowably wrong. Parent/component/history and unknown identity remain inspectable without fabricated matches or erasure of superseded evidence.",
            "F": "Current register, split/merge/alias/correction traces, exceptions, claim-count reconciliation, competing assertions and subsequent-events note coherently fulfil both historical views and follow-through.",
            "M": "Same percentage/different membership is the prominent argument. Side-by-side sets and changed-unit filter prevent an unchanged66.7% headline from concealing the correction; source detail and action queue are easy to reach.",
            "B": "Coordinator browser showed the unchanged8/12 with W102/W110 swap, U-99 zero-match and OLD-17 trace. Independent DOM check verified both sets, all six traces, changed filter, source targets and reset. No numerical free input or unsupported identity edit is exposed."
          },
          "material_failures": [],
          "findings": [],
          "verification_gaps": [
            "This is a separate AI assessment within the same model family, not Claude or human professional validation.",
            "Native browser evidence is the coordinator’s bounded ordinary-browser observations; the assessor independently ran JSDOM checks, which do not establish comprehensive rendering, keyboard or assistive-technology coverage."
          ],
          "disposition": "verification only",
          "independent_checks": [
            "Authorised denominator12−1+2−2+1=12; current claim coverage16−1duplicate−1parent−1unresolved−2merge consolidation=11.",
            "As-known accepted set W101,W102,W103,W104A,W105,W106,W1089,W111; restated removesW102/addsW110. DOM sets agree.",
            "W102 distance5m; W107 distance30m; later W112 event excluded. Six lineage disclosures and unresolved search tested.",
            "Reviewer DOM outcomes: 13 passed checks, 0 failed assertions. See saved checking code and results; these are not the author’s self-tests."
          ],
          "artifact_sha256": "72f2dedaa839cc1276a998b3676a9982c2fc79babe661a831e0e07a0fa0b6b7d",
          "status": "provisional AI assessment",
          "rationale": "AI review disposition: verification only",
          "served_artifact_sha256": "6972b4615f811360aefaa55dedafc594f42679bd33e270dde795c049a49b0503",
          "review_evidence_id": "essai03-separate-review-08-14"
        }
      },
      "status": "first attempt retained"
    },
    {
      "id": "BPB2-11",
      "self": {
        "initial": {
          "reviewer": "gpt-6-astra · answering agent",
          "method": "Unverified self-assessment before feedback; common assessment/source packaging added afterward",
          "assessed_at": "2026-11-14T18:36:36.955165Z",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 3,
            "B": 3
          },
          "reasons": {
            "C": "Reconciles £12.95m base without accrual double counting; schedules all six option/outcome combinations under working-day and building-empty rules; identifies next-working-day grant failure and authority shortfall. Fourteen local DOM/calculation checks pass.",
            "D": "Provides a conditional preferred strategy, an affordable fallback within the supplied variation range, explicit dated opening meaning, permanent reserve, early cash plan, and decision-reversal thresholds.",
            "E": "All eight original records remain inspectable; material claims cite source IDs. Probabilities, variation range, disputed recovery, future acceptance and donor entitlement remain explicitly uncertain. Lower-variation cash timing assumption is disclosed.",
            "F": "Includes four panels, all four strategies, six outcomes, probability/variation controls, occupation calendar, dated sequence, cost/cash trail, reset, source recovery and gallery link.",
            "M": "Uses an outcome-led board briefing with distinctive decision, occupation, funding and challenge panels. Dense comparison and funding content may require scrolling; no rendered-layout inspection was available.",
            "B": "JSDOM executes every control and all 14 checks pass, including invalid input, reset, reload and evidence drawer. No ordinary browser rendering, keyboard or viewport check was run."
          },
          "material_failures": [],
          "rationale": "Self-assessment only.",
          "artifact_sha256": "98b1d6bca616a70f07dbbecece72a3f641916480f49467cc76659af012927846",
          "verification_gaps": []
        },
        "current": {
          "reviewer": "gpt-6-astra · answering agent",
          "method": "Unverified self-assessment after logged assisted repair cycle 1; first-attempt score is retained separately",
          "assessed_at": "2026-11-14T18:50:15.220460Z",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 3,
            "B": 3
          },
          "reasons": {
            "C": "Strategy identifiers are normalized before delivery weighting. The two confirmed contradictions fail before repair and pass afterward. All 14 original checks and 8 new regression groups pass, including 140 cross-strategy combinations plus 137 baseline-counterpart probability/variation sweeps.",
            "D": "Preferred-strategy expected funding in the visible decision summary now agrees with the IV comparison row and detailed expected metric. Late stress, authority gap, conditional approval and affordable fallback remain separate and unchanged.",
            "E": "All eight original records remain inspectable; material claims cite source IDs. Probabilities, variation range, disputed recovery, future acceptance and donor entitlement remain explicitly uncertain. Lower-variation cash timing assumption is disclosed.",
            "F": "Includes four panels, all four strategies, six outcomes, probability/variation controls, occupation calendar, dated sequence, cost/cash trail, reset, source recovery and gallery link.",
            "M": "Uses an outcome-led board briefing with distinctive decision, occupation, funding and challenge panels. Dense comparison and funding content may require scrolling; no rendered-layout inspection was available.",
            "B": "22 local DOM/calculation test groups pass. New checks read visible decision-summary numbers and compare them with independently calculated expectations and detailed results. No rendered-browser, viewport or keyboard check was run by this repair agent."
          },
          "material_failures": [],
          "rationale": "Assisted repair: corrected the expected-funding summary propagation. Original first answer remains unchanged.",
          "artifact_sha256": "8329b9f75487de996fc84528d3169b85e85da2f5ab0b26f1cae1d945c25fb248",
          "served_artifact_sha256": "8f38bfcafae4d67be4086e46f89e205c46e80ac0777ca0feb1ad7c7290ee6ba9",
          "repair_cycle": 1,
          "verification_gaps": []
        }
      },
      "checked": {
        "initial": {
          "reviewer": "gpt-6-astra · separate assessment agent",
          "method": "Separate source/logic review, independent checks and coordinator ordinary-browser observations; same model family, not a human or Claude assessment",
          "assessed_at": "2026-11-14T18:49:43.095990Z",
          "case_id": "BPB2-11",
          "dimensions": {
            "C": 2,
            "D": 3,
            "E": 4,
            "F": 4,
            "M": 3,
            "B": 2
          },
          "reasons": {
            "C": "Core baseline, six strategy/outcome costs, activity dates and late funding gate are correct, but the required probability/variation interaction displays materially wrong expected funding in the decision panel. At100%/.45, it says £1.260m expected while the correct value and its own finance metric are £0.220m. At70%/.60 it says £1.410m rather than £0.682m.",
            "D": "The baseline conditional expedition/staged recommendation, funding reserve, £0.49 variation gate, normal-staged fallback and decision owners are usable. Decision usefulness falls below4 because its changed-scenario recommendation paragraph gives the late amount as the expected requirement, distorting the comparison a governor is asked to use.",
            "E": "Physical80.5% is validly distinguished from readiness; the gross ledger, prospective grant, unsigned future acceptance, disputed designer recovery and temporary bridge are transparently separated. The high variation is not given invented probability and the unresolved authority gap is retained.",
            "F": "All four visual panels and all supplied occupation strategies are present, with dated sequence, 7June experience, gross/expected costs, permanent funding, cash timing, acceptance route and switch conditions. The defect is incorrect propagation within an existing required element, not an omitted workstream.",
            "M": "Clear panel structure and dates/units, but the first panel is dense and its sensitivity sentence contradicts the displayed finance metric after ordinary scenario changes. A governor can inspect the correct workings, but must resolve a contradiction in the main decision text.",
            "B": "Controls, evidence, reset and invalid-variation handling work in targeted coordinator/browser and DOM observations. Required dependency propagation materially fails: the decision-panel expected funding does not follow the probability correctly, despite correct panel3 values. This is the same defect counted in C."
          },
          "material_failures": [
            {
              "code": "MF-C",
              "description": "Probability/variation changes overstate preferred strategy IV expected permanent funding in the main decision panel by substituting the late outcome, while the finance metric remains correct.",
              "source": "BPB2-11 retained first index.html:58,62,64; #decision-sensitivity versus #expected-gap; 11-F probability and 11-G funding rules; independent-dom-results.json case11.",
              "consequence": "At100%/.45,£1.260m is presented as expected rather than£0.220m; at70%/.60,£1.410m rather than£0.682m. This is a consequential funding statement a governor could use; baseline and late reserve remain correct.",
              "status": "open"
            },
            {
              "code": "MF-I",
              "description": "Required probability/variation propagation reaches the finance panel but fails in the recommendation text.",
              "source": "BPB2-11 retained first index.html:58,62,64; #decision-sensitivity versus #expected-gap; 11-F probability and 11-G funding rules; independent-dom-results.json case11.",
              "consequence": "Conflicting displayed outputs after ordinary required interaction; B2 shares the same underlying defect as MF-C, not a second independent arithmetic error.",
              "status": "open"
            }
          ],
          "findings": [
            {
              "severity": "material",
              "description": "Probability/variation changes overstate preferred strategy IV expected permanent funding in the main decision panel by substituting the late outcome, while the finance metric remains correct.",
              "evidence": "Reproduced8 wrong decision sentences among12 probability/variation states; raw function weighted expects an object but is called with a string for preferredIV.",
              "proposed_repair": "In a separate repaired copy, normalise strategy IDs inside weighted or pass the strategy object; use one consistent expected-value result in both panels. Recheck baseline and p0/15/70/100 across V.25/.45/.60, including native ordinary-browser comparison."
            },
            {
              "severity": "minor",
              "description": "The main decision panel is dense despite the requested four-panel board format.",
              "evidence": "Four visual panels exist, but panel1 carries recommendation, cost gate, authority, calendar, detailed conditional resolution and sensitivity narrative.",
              "proposed_repair": "Optionally shorten panel1 to decision, three gates and fallback; retain detailed resolution in its disclosure. This is not required to fix the material calculation."
            },
            {
              "severity": "observation",
              "description": "The conditional strategy is defensible: it explicitly withholds expedition if the late/high-variation exposure cannot be brought within authority and gives the affordable normal-staged fallback. The candidate key does not prescribe a unique strategy.",
              "evidence": "Source pack, answer and candidate key compared after independent interpretation; see independent-interpretation.md and independent checks.",
              "proposed_repair": "No repair required on this point; retain the defensible alternative rather than force key wording or allocation."
            }
          ],
          "verification_gaps": [
            "This is a separate AI assessment within the same model family, not Claude or human professional validation.",
            "Native browser evidence is the coordinator’s bounded ordinary-browser observations; the assessor independently ran JSDOM checks, which do not establish comprehensive rendering, keyboard or assistive-technology coverage.",
            "Targeted native confirmation of the newly found decision-panel inconsistency at100%/.45 and70%/.60 was requested from the coordinator; first score does not depend on repairing the artifact."
          ],
          "disposition": "usable after bounded repair",
          "independent_checks": [
            "Independently rebuilt dry/wet tests, training and validation: full occupation14June expedited-on-time,21June normal/late; staged closures10–11June only in normal/late.",
            "24 variation/strategy/outcome combinations agreed for dates, temporary days, gross cost, grant and funding;12 probability/variation expected-value scenarios checked.",
            "Eight non-baseline scenario sentences were wrong; at p100,V.45 correct £0.220m versus £1.260m decision text; at p70,V.60 correct £0.682m versus £1.410m.",
            "Underlying cause: raw index.html line58 weighted(s) reads s.exp; line62 passes the string exp-staged; line64 renders rec.gap. Baseline hard-coded decision sentence masks the issue.",
            "Reviewer DOM outcomes: 32 passed checks, 8 failed assertions. See saved checking code and results; these are not the author’s self-tests."
          ],
          "artifact_sha256": "98b1d6bca616a70f07dbbecece72a3f641916480f49467cc76659af012927846",
          "status": "provisional AI assessment",
          "rationale": "AI review disposition: usable after bounded repair",
          "served_artifact_sha256": "6198d9dec78003d2ed1c7c5acc547c04f5595b6bcc33b1685d26026568c2980a",
          "review_evidence_id": "essai03-separate-review-08-14"
        },
        "current": {
          "case_id": "BPB2-11",
          "reviewer": "gpt-6-astra · separate assessment agent",
          "method": "Separate source/logic review, independent checks and coordinator ordinary-browser observations; same model family, not a human or Claude assessment. Assisted repair cycle1 recheck, not a first attempt.",
          "assessed_at": "2026-11-14T18:54:26.646072Z",
          "stage": "assisted repair cycle1",
          "assisted_repair_cycle": 1,
          "assistance_status": "Reviewer supplied exact defect, examples and root cause; original author repaired a separate copy. First-attempt rating is preserved unchanged.",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 3,
            "B": 4
          },
          "reasons": {
            "C": "The wrong expected-funding sentence is corrected. Independent checks now agree across120 selected-strategy/probability/variation states, the two reported scenarios and24 original dated outcomes. Costs, grant, cash, temporary-teaching days and adverse funding gate are unchanged and correct; expected versus late-state funding reconciles visibly.",
            "D": "The conditional expedition/staged recommendation now carries a correct expected funding comparison, owners, £0.49m variation gate and normal-staged fallback. Late/high exposure remains unauthorised at£1.41m and is not erased by the repair; decision-switch conditions remain clear.",
            "E": "Original evidence and uncertainty treatment are byte-for-byte unchanged outside the corrected calculation function. Physical readiness, gross obligations, disputed recovery, prospective grant, bridge, variation uncertainty and future sign-offs remain distinct; no new authority or evidence is invented.",
            "F": "All four panels, four strategies, dated outcomes, cash timing, funding, conditions, fallback and inspectable evidence remain present. The repair is exactly one calculation-function change; no required capability or source was removed.",
            "M": "The visible contradiction has gone and the four-panel structure remains clear, but the unchanged first decision panel still combines a lengthy recommendation, funding gate, authority/date conditions and sensitivity narrative. It remains clear enough for the named user but falls short of concise prioritisation for4.",
            "B": "Coordinator ordinary-browser recheck confirms both corrected scenario amounts agree across decision and finance panels. Independent DOM checks cover cross-surface propagation, all selected strategies, invalid variation retention, exact funding boundary and reset. Original native navigation/outcome evidence remains applicable because the sole HTML change is weighted() normalisation. No material required interaction defect remains in tested paths."
          },
          "material_failures": [
            {
              "code": "MF-C",
              "description": "Probability/variation changes overstate preferred strategy IV expected permanent funding in the main decision panel by substituting the late outcome, while the finance metric remains correct.",
              "source": "BPB2-11 retained first index.html:58,62,64; #decision-sensitivity versus #expected-gap; 11-F probability and 11-G funding rules; independent-dom-results.json case11.",
              "consequence": "At100%/.45,£1.260m is presented as expected rather than£0.220m; at70%/.60,£1.410m rather than£0.682m. This is a consequential funding statement a governor could use; baseline and late reserve remain correct.",
              "status": "resolved",
              "resolution": "Resolved in assisted repair cycle1 only. Independent DOM and coordinator ordinary-browser checks confirm the decision summary and finance expected values agree: £0.220m at100%/.45 and £0.682m at70%/.60, with late amounts £1.260m/£1.410m separately retained. The original first-attempt defect remains present in its frozen artifact.",
              "resolution_evidence": "independent-recheck-results.json (158 passed,0 failed); coordinator-browser-observations.json stage current-repair01 at2026-11-14T18:52:34.441Z; first-to-repair.diff."
            },
            {
              "code": "MF-I",
              "description": "Required probability/variation propagation reaches the finance panel but fails in the recommendation text.",
              "source": "BPB2-11 retained first index.html:58,62,64; #decision-sensitivity versus #expected-gap; 11-F probability and 11-G funding rules; independent-dom-results.json case11.",
              "consequence": "Conflicting displayed outputs after ordinary required interaction; B2 shares the same underlying defect as MF-C, not a second independent arithmetic error.",
              "status": "resolved",
              "resolution": "Resolved in assisted repair cycle1 only. Independent DOM and coordinator ordinary-browser checks confirm the decision summary and finance expected values agree: £0.220m at100%/.45 and £0.682m at70%/.60, with late amounts £1.260m/£1.410m separately retained. The original first-attempt defect remains present in its frozen artifact.",
              "resolution_evidence": "independent-recheck-results.json (158 passed,0 failed); coordinator-browser-observations.json stage current-repair01 at2026-11-14T18:52:34.441Z; first-to-repair.diff."
            }
          ],
          "findings": [
            {
              "severity": "observation",
              "description": "The original MF-C/MF-I share one calculation cause and are both resolved in the assisted copy only.",
              "evidence": "At100%/.45 both surfaces show£0.220m expected, late£1.260m; at70%/.60 both show£0.682m expected, late£1.410m. Native coordinator evidence corroborates. Frozen first hash and its original contradiction are retained.",
              "proposed_repair": "No further correction required for this defect. Keep first and assisted ratings/artifacts separate."
            },
            {
              "severity": "minor",
              "description": "The main decision panel remains dense for a four-panel board briefing.",
              "evidence": "The repair only changes the weighted calculation; the original panel text and disclosure layout are unchanged.",
              "proposed_repair": "Optional: shorten the default decision panel to the resolution, funding gate and fallback, retaining detail in its disclosure. This is not needed for the readiness gate."
            }
          ],
          "verification_gaps": [
            "This is same-model-family AI rechecking, not human or Claude professional validation.",
            "Native confirmation comes from the coordinator’s targeted ordinary-browser observations; the reviewer independently used JSDOM, not CUA. Comprehensive keyboard, assistive-technology and viewport certification remains unperformed.",
            "Author regression outputs were inspected as self-test evidence; the158 reviewer checks are separate. No adaptation, workplace outcome or professional time saving is inferred."
          ],
          "disposition": "verification only",
          "independent_checks": [
            "Read unchanged case.txt, REPAIR.md and author regression outputs; verified case bytes equal the frozen model-visible case. Did not read author self-score files.",
            "Compared raw first/repaired HTML: only weighted() changes, normalising string strategy IDs before reading the expedition flag. All remaining HTML is identical.",
            "Recomputed100%/.45: expected£0.220m, late£1.260m;70%/.60: expected£0.682m, late£1.410m. Both required cases now agree in visible summary/finance.",
            "120 combinations of six probabilities, five variations and four selected strategies: checked preferred-IV summary, selected metrics, all four option rows and string/object argument equivalence.",
            "24 independently derived original strategy/variation/delivery outcomes retain dates, temporary-teaching mix, cost, grant and permanent gap.",
            "Checked exact£0.49/£0.50 funding boundary, five invalid variations retaining last valid values, fragment targets, no script errors and baseline/reset£0.532m expected/£1.260m late.",
            "158 checks passed; zero failed. Reopened frozen first in separate DOM and reproduced its original70%/.60 contradiction; original SHA-256 unchanged."
          ],
          "artifact_sha256": "8329b9f75487de996fc84528d3169b85e85da2f5ab0b26f1cae1d945c25fb248",
          "original_artifact_sha256": "98b1d6bca616a70f07dbbecece72a3f641916480f49467cc76659af012927846",
          "original_dimensions": {
            "C": 2,
            "D": 3,
            "E": 4,
            "F": 4,
            "M": 3,
            "B": 2
          },
          "original_pm_score": 76.25,
          "current_pm_score": 97.5,
          "status": "provisional AI assessment after assisted repair",
          "served_artifact_sha256": "8f38bfcafae4d67be4086e46f89e205c46e80ac0777ca0feb1ad7c7290ee6ba9",
          "repair_cycle": 1,
          "review_evidence_id": "essai03-separate-recheck11-cycle01",
          "rationale": "Assisted repair cycle1 independently rechecked by the same separate assessor; original first-attempt ratings remain unchanged."
        }
      },
      "status": "first attempt and assisted repair retained",
      "repair_cycles": 1
    },
    {
      "id": "BPB2-12",
      "self": {
        "initial": {
          "reviewer": "gpt-6-astra · answering agent",
          "method": "Unverified self-assessment before feedback; common assessment/source packaging added afterward",
          "assessed_at": "2026-11-14T17:58:35.186318+00:00",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 3,
            "B": 3
          },
          "reasons": {
            "C": "All twelve base arrangement/week combinations reconcile 120 appointments, every cohort and permitted mode/window/role, all 42 session pools, remote trigger/cap, onboarding, vaccination exception, mandatory concession offsets, weekly admin and six-week authority. A separately authored Python calculation agrees with the page.",
            "D": "Recommends the stable £2,700 package against the £1,800 alternative; quantifies premium, access, role continuity, displacement, approval terms, and both booking and resource contingencies.",
            "E": "Source IDs accompany material claims; full supplied evidence remains readable; conditional offers, individual acceptance and delivery are separate; no future uptake or approval is invented. Cost withdrawal is explicitly pre-contract and creates no refund assumption.",
            "F": "Two complete six-week arrangements and lower-cost middle-week variation; all concession sign-offs, proposed deadlines, launch/monitoring/exit, allocation rule, fallback and required controls included.",
            "M": "Clear recommendation and layered calculations, but the full negotiation and evidence page is long and has dense inspectable resource tables.",
            "B": "JSDOM checks passed all control changes, reset, generated tables and resource outcomes with no JavaScript errors. Actual browser rendering has not been checked, so no highest browser score is claimed."
          },
          "material_failures": [],
          "rationale": "Self-assessment only.",
          "artifact_sha256": "6cf97242a865f9b14091efc874e3e326de9b231a7f9d9ea2340c993f76f3ec1b",
          "verification_gaps": []
        },
        "current": {
          "reviewer": "gpt-6-astra · answering agent",
          "method": "Unverified self-assessment before feedback; common assessment/source packaging added afterward; current case analysis unchanged",
          "assessed_at": "2026-11-14T17:58:35.186318+00:00",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 3,
            "B": 3
          },
          "reasons": {
            "C": "All twelve base arrangement/week combinations reconcile 120 appointments, every cohort and permitted mode/window/role, all 42 session pools, remote trigger/cap, onboarding, vaccination exception, mandatory concession offsets, weekly admin and six-week authority. A separately authored Python calculation agrees with the page.",
            "D": "Recommends the stable £2,700 package against the £1,800 alternative; quantifies premium, access, role continuity, displacement, approval terms, and both booking and resource contingencies.",
            "E": "Source IDs accompany material claims; full supplied evidence remains readable; conditional offers, individual acceptance and delivery are separate; no future uptake or approval is invented. Cost withdrawal is explicitly pre-contract and creates no refund assumption.",
            "F": "Two complete six-week arrangements and lower-cost middle-week variation; all concession sign-offs, proposed deadlines, launch/monitoring/exit, allocation rule, fallback and required controls included.",
            "M": "Clear recommendation and layered calculations, but the full negotiation and evidence page is long and has dense inspectable resource tables.",
            "B": "JSDOM checks passed all control changes, reset, generated tables and resource outcomes with no JavaScript errors. Actual browser rendering has not been checked, so no highest browser score is claimed."
          },
          "material_failures": [],
          "rationale": "Self-assessment only.",
          "artifact_sha256": "6cf97242a865f9b14091efc874e3e326de9b231a7f9d9ea2340c993f76f3ec1b",
          "verification_gaps": []
        }
      },
      "checked": {
        "initial": {
          "reviewer": "gpt-6-astra · separate assessment agent",
          "method": "Separate source/logic review, independent checks and coordinator ordinary-browser observations; same model family, not a human or Claude assessment",
          "assessed_at": "2026-11-14T18:49:43.095990Z",
          "case_id": "BPB2-12",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 4,
            "B": 4
          },
          "reasons": {
            "C": "Independently resummed every cohort, legal session/mode/role, all42 session resource cells, weekly admin, overheads, review, remote ceiling and offer offsets for all12 baseline package/week states plus30 withdrawals. All agreed. Both complete six-week proposals fit; no averaged onboarding or W3–4 capacity is used.",
            "D": "Recommends £2,700 P+R for stable operation and explains the £900 premium over T+V+R, exact displaced work, narrower alternative headroom and authority reserve. Quantifies both P withdrawal and a declined video booking, with bounded decisions instead of unsupported automatic rebooking.",
            "E": "Planning permission, offers, patient acceptance and delivered care are distinct. Jo does not supply consent; prioritisation is proposed for Dr Shah approval. Conditional concessions, deferred teaching liability and actual unrecorded uptake remain explicit and source-linked.",
            "F": "Two complete arrangements, middle-week changes, all concessions/offsets, six-week cost, approval terms/deadlines, transparent allocation principle, launch/digest monitoring and expiry are covered. Exact resource and reduced-service contingencies connect the technical allocation to negotiation.",
            "M": "The proposed stable service and practical premium lead, with a side-by-side comparison and selected-week matrix. Detailed slots, clinician roles and overheads remain inspectable; the distinction among planned/offered/accepted/delivered is explicit.",
            "B": "Coordinator browser changed arrangement/week and withdrew T, showing invalidity. Independent DOM exercised42 full states, checked each displayed resource cell against independent arithmetic, costs/overheads and reset. Withdrawals preserve allocation demand and honestly invalidate the fixed template."
          },
          "material_failures": [],
          "findings": [
            {
              "severity": "observation",
              "description": "The alternative allocation differs from candidate-key witness:33 remote and1 minute onboarding slack rather than32 remote and3 minutes slack. Independent per-session checks prove it feasible; exact key allocations are not required.",
              "evidence": "Source pack, answer and candidate key compared after independent interpretation; see independent-interpretation.md and independent checks.",
              "proposed_repair": "No repair required on this point; retain the defensible alternative rather than force key wording or allocation."
            }
          ],
          "verification_gaps": [
            "This is a separate AI assessment within the same model family, not Claude or human professional validation.",
            "Native browser evidence is the coordinator’s bounded ordinary-browser observations; the assessor independently ran JSDOM checks, which do not establish comprehensive rendering, keyboard or assistive-technology coverage."
          ],
          "disposition": "verification only",
          "independent_checks": [
            "Recomputed12 baseline states and30 single-concession withdrawals from allocation data using independently encoded fixture rules; all42 matrices agree with page.",
            "P+R:20remote,admin1042 W1–2 and1002 later; alternative33remote normal/36middle,admin1084/1050/1044 byweektype. All120cohorts reconcile.",
            "P withdrawal leaves Friday shortages120consultant/30nurse/120room; six SV consultant physical slots exactly account for them. Wednesday BV video→physical adds20room/5nurse, exceeding room cap.",
            "Reviewer DOM outcomes: 45 passed checks, 0 failed assertions. See saved checking code and results; these are not the author’s self-tests."
          ],
          "artifact_sha256": "6cf97242a865f9b14091efc874e3e326de9b231a7f9d9ea2340c993f76f3ec1b",
          "status": "provisional AI assessment",
          "rationale": "AI review disposition: verification only",
          "served_artifact_sha256": "dd31917e0001d0be98c85cc7673e14e269bf7ba7d96109dfe3c674476f0ad414",
          "review_evidence_id": "essai03-separate-review-08-14"
        },
        "current": {
          "reviewer": "gpt-6-astra · separate assessment agent",
          "method": "Separate source/logic review, independent checks and coordinator ordinary-browser observations; same model family, not a human or Claude assessment; current case analysis is unchanged from the retained first answer",
          "assessed_at": "2026-11-14T18:49:43.095990Z",
          "case_id": "BPB2-12",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 4,
            "B": 4
          },
          "reasons": {
            "C": "Independently resummed every cohort, legal session/mode/role, all42 session resource cells, weekly admin, overheads, review, remote ceiling and offer offsets for all12 baseline package/week states plus30 withdrawals. All agreed. Both complete six-week proposals fit; no averaged onboarding or W3–4 capacity is used.",
            "D": "Recommends £2,700 P+R for stable operation and explains the £900 premium over T+V+R, exact displaced work, narrower alternative headroom and authority reserve. Quantifies both P withdrawal and a declined video booking, with bounded decisions instead of unsupported automatic rebooking.",
            "E": "Planning permission, offers, patient acceptance and delivered care are distinct. Jo does not supply consent; prioritisation is proposed for Dr Shah approval. Conditional concessions, deferred teaching liability and actual unrecorded uptake remain explicit and source-linked.",
            "F": "Two complete arrangements, middle-week changes, all concessions/offsets, six-week cost, approval terms/deadlines, transparent allocation principle, launch/digest monitoring and expiry are covered. Exact resource and reduced-service contingencies connect the technical allocation to negotiation.",
            "M": "The proposed stable service and practical premium lead, with a side-by-side comparison and selected-week matrix. Detailed slots, clinician roles and overheads remain inspectable; the distinction among planned/offered/accepted/delivered is explicit.",
            "B": "Coordinator browser changed arrangement/week and withdrew T, showing invalidity. Independent DOM exercised42 full states, checked each displayed resource cell against independent arithmetic, costs/overheads and reset. Withdrawals preserve allocation demand and honestly invalidate the fixed template."
          },
          "material_failures": [],
          "findings": [
            {
              "severity": "observation",
              "description": "The alternative allocation differs from candidate-key witness:33 remote and1 minute onboarding slack rather than32 remote and3 minutes slack. Independent per-session checks prove it feasible; exact key allocations are not required.",
              "evidence": "Source pack, answer and candidate key compared after independent interpretation; see independent-interpretation.md and independent checks.",
              "proposed_repair": "No repair required on this point; retain the defensible alternative rather than force key wording or allocation."
            }
          ],
          "verification_gaps": [
            "This is a separate AI assessment within the same model family, not Claude or human professional validation.",
            "Native browser evidence is the coordinator’s bounded ordinary-browser observations; the assessor independently ran JSDOM checks, which do not establish comprehensive rendering, keyboard or assistive-technology coverage."
          ],
          "disposition": "verification only",
          "independent_checks": [
            "Recomputed12 baseline states and30 single-concession withdrawals from allocation data using independently encoded fixture rules; all42 matrices agree with page.",
            "P+R:20remote,admin1042 W1–2 and1002 later; alternative33remote normal/36middle,admin1084/1050/1044 byweektype. All120cohorts reconcile.",
            "P withdrawal leaves Friday shortages120consultant/30nurse/120room; six SV consultant physical slots exactly account for them. Wednesday BV video→physical adds20room/5nurse, exceeding room cap.",
            "Reviewer DOM outcomes: 45 passed checks, 0 failed assertions. See saved checking code and results; these are not the author’s self-tests."
          ],
          "artifact_sha256": "6cf97242a865f9b14091efc874e3e326de9b231a7f9d9ea2340c993f76f3ec1b",
          "status": "provisional AI assessment",
          "rationale": "AI review disposition: verification only",
          "served_artifact_sha256": "a0c00b43e0ec953bf0296254935ccc1c813dd7830406add815e56e382c2ef4f2",
          "review_evidence_id": "essai03-separate-review-08-14"
        }
      },
      "status": "first attempt retained"
    },
    {
      "id": "BPB2-13",
      "self": {
        "initial": {
          "reviewer": "gpt-6-astra · answering agent",
          "method": "Unverified self-assessment before feedback; common assessment/source packaging added afterward",
          "assessed_at": "2026-11-14T18:36:30Z",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 3,
            "B": 3
          },
          "reasons": {
            "C": "All three valid legacy records are retained; the other three static families and both campaign dynamics are scheduled. Whole-day dependencies, P3 acceptance, two distinct permissions and every supplied shared-resource/window restriction are respected. A separately written bounded search confirms D9-D10 infeasibility with one crew and feasibility with two.",
            "D": "The page leads with the £8,000 delegated crew decision and proposed coordinator appointment, separates sponsor decisions, compares D13/£8,000 with D18/£12,000, and states D12 is unattainable in both routes.",
            "E": "Source IDs accompany material claims and all supplied records remain readable. Signed historical passes, derived applicability, future results, prospective approvals and hypothetical control changes are distinguished.",
            "F": "Provides eight selectable test entries, 17 detailed critical handoffs, linked day-by-day resources, both crew scenarios, alternate zone order, reset, full source recovery and gallery return.",
            "M": "Practical recommendation, comparison and permission chain are concise. The complete interface register is intentionally detailed and wide; actual browser visual quality has not been inspected.",
            "B": "Eleven JSDOM self-check groups pass, including interaction changes, source links, reset and four resource/dependency combinations. No real browser rendering or ordinary UI verification was available to this solver."
          },
          "material_failures": [],
          "rationale": "Self-assessment only.",
          "artifact_sha256": "d8c7f69530875f185efc88319697eff8e062dedd8dbb403cc5b62ed97fda0ef7",
          "verification_gaps": []
        },
        "current": {
          "reviewer": "gpt-6-astra · answering agent",
          "method": "Unverified self-assessment before feedback; common assessment/source packaging added afterward; current case analysis unchanged",
          "assessed_at": "2026-11-14T18:36:30Z",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 3,
            "B": 3
          },
          "reasons": {
            "C": "All three valid legacy records are retained; the other three static families and both campaign dynamics are scheduled. Whole-day dependencies, P3 acceptance, two distinct permissions and every supplied shared-resource/window restriction are respected. A separately written bounded search confirms D9-D10 infeasibility with one crew and feasibility with two.",
            "D": "The page leads with the £8,000 delegated crew decision and proposed coordinator appointment, separates sponsor decisions, compares D13/£8,000 with D18/£12,000, and states D12 is unattainable in both routes.",
            "E": "Source IDs accompany material claims and all supplied records remain readable. Signed historical passes, derived applicability, future results, prospective approvals and hypothetical control changes are distinguished.",
            "F": "Provides eight selectable test entries, 17 detailed critical handoffs, linked day-by-day resources, both crew scenarios, alternate zone order, reset, full source recovery and gallery return.",
            "M": "Practical recommendation, comparison and permission chain are concise. The complete interface register is intentionally detailed and wide; actual browser visual quality has not been inspected.",
            "B": "Eleven JSDOM self-check groups pass, including interaction changes, source links, reset and four resource/dependency combinations. No real browser rendering or ordinary UI verification was available to this solver."
          },
          "material_failures": [],
          "rationale": "Self-assessment only.",
          "artifact_sha256": "d8c7f69530875f185efc88319697eff8e062dedd8dbb403cc5b62ed97fda0ef7",
          "verification_gaps": []
        }
      },
      "checked": {
        "initial": {
          "reviewer": "gpt-6-astra · separate assessment agent",
          "method": "Separate source/logic review, independent checks and coordinator ordinary-browser observations; same model family, not a human or Claude assessment",
          "assessed_at": "2026-11-14T18:49:43.095990Z",
          "case_id": "BPB2-13",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 4,
            "B": 4
          },
          "reasons": {
            "C": "Independently proved retained E-G/E-R/W-R and new W-G/E-P/W-P, earliest static/energisation/familiarisation sequence and operator conflict. Both zone orders give serviceD13/20October with £8,000 extra crew and D18/27October with £12,000 rebooking; neither meetsD12.",
            "D": "Clear immediate £8,000 delegated crew purchase/coordinator appointment decision, explicit sponsor date change, and cheaper/five-day-earlier trade-off. No-crew fallback and its sponsor rebooking authority are fully stated.",
            "E": "Uses relevant family fingerprints rather than executable label alone; passed source evidence remains valid where appropriate. TTA is derived from the signed charter and does not become passenger permission; proposed coordinator and future successes/signatures remain unapproved.",
            "F": "Critical interface register names producer, receiving authority, controlled version, evidence, accountable decision and due point. It joins input release, QA, TTA, energisation/familiarisation, dynamic/cross-zone and final service record to a resource-feasible daily view.",
            "M": "Immediate decision and two-route comparison lead; concise proof explains why one crew cannot makeD9. Clickable test/interface detail and day links make the large register inspectable without obscuring the service-date consequence.",
            "B": "Coordinator browser checked crew/West-first switch, test disclosure, TTA authority and linked day return. Independent checks verified four route states against exhaustive operator-slot feasibility, all eight test disclosures, fragment targets and reset."
          },
          "material_failures": [],
          "findings": [],
          "verification_gaps": [
            "This is a separate AI assessment within the same model family, not Claude or human professional validation.",
            "Native browser evidence is the coordinator’s bounded ordinary-browser observations; the assessor independently ran JSDOM checks, which do not establish comprehensive rendering, keyboard or assistive-technology coverage."
          ],
          "disposition": "verification only",
          "independent_checks": [
            "Enumerated familiarisation timing against fixed inseparable dynamic windows and available operator crews; earliest feasible windows9–10 with extra and14–15 without.",
            "Checked both first-zone choices, no operator overlap beyond availablecrewcount, final assemblyD12/D17 and passenger startD13/D18.",
            "£8,000 crew versus£12,000 rebook; five working days and£4,000 difference. All static/dynamic applicability disclosures tested.",
            "Reviewer DOM outcomes: 15 passed checks, 0 failed assertions. See saved checking code and results; these are not the author’s self-tests."
          ],
          "artifact_sha256": "d8c7f69530875f185efc88319697eff8e062dedd8dbb403cc5b62ed97fda0ef7",
          "status": "provisional AI assessment",
          "rationale": "AI review disposition: verification only",
          "served_artifact_sha256": "c09c465f05f5348051d15ab3f214a82eb220632c6930bdde9d43f375cb0005cb",
          "review_evidence_id": "essai03-separate-review-08-14"
        },
        "current": {
          "reviewer": "gpt-6-astra · separate assessment agent",
          "method": "Separate source/logic review, independent checks and coordinator ordinary-browser observations; same model family, not a human or Claude assessment; current case analysis is unchanged from the retained first answer",
          "assessed_at": "2026-11-14T18:49:43.095990Z",
          "case_id": "BPB2-13",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 4,
            "B": 4
          },
          "reasons": {
            "C": "Independently proved retained E-G/E-R/W-R and new W-G/E-P/W-P, earliest static/energisation/familiarisation sequence and operator conflict. Both zone orders give serviceD13/20October with £8,000 extra crew and D18/27October with £12,000 rebooking; neither meetsD12.",
            "D": "Clear immediate £8,000 delegated crew purchase/coordinator appointment decision, explicit sponsor date change, and cheaper/five-day-earlier trade-off. No-crew fallback and its sponsor rebooking authority are fully stated.",
            "E": "Uses relevant family fingerprints rather than executable label alone; passed source evidence remains valid where appropriate. TTA is derived from the signed charter and does not become passenger permission; proposed coordinator and future successes/signatures remain unapproved.",
            "F": "Critical interface register names producer, receiving authority, controlled version, evidence, accountable decision and due point. It joins input release, QA, TTA, energisation/familiarisation, dynamic/cross-zone and final service record to a resource-feasible daily view.",
            "M": "Immediate decision and two-route comparison lead; concise proof explains why one crew cannot makeD9. Clickable test/interface detail and day links make the large register inspectable without obscuring the service-date consequence.",
            "B": "Coordinator browser checked crew/West-first switch, test disclosure, TTA authority and linked day return. Independent checks verified four route states against exhaustive operator-slot feasibility, all eight test disclosures, fragment targets and reset."
          },
          "material_failures": [],
          "findings": [],
          "verification_gaps": [
            "This is a separate AI assessment within the same model family, not Claude or human professional validation.",
            "Native browser evidence is the coordinator’s bounded ordinary-browser observations; the assessor independently ran JSDOM checks, which do not establish comprehensive rendering, keyboard or assistive-technology coverage."
          ],
          "disposition": "verification only",
          "independent_checks": [
            "Enumerated familiarisation timing against fixed inseparable dynamic windows and available operator crews; earliest feasible windows9–10 with extra and14–15 without.",
            "Checked both first-zone choices, no operator overlap beyond availablecrewcount, final assemblyD12/D17 and passenger startD13/D18.",
            "£8,000 crew versus£12,000 rebook; five working days and£4,000 difference. All static/dynamic applicability disclosures tested.",
            "Reviewer DOM outcomes: 15 passed checks, 0 failed assertions. See saved checking code and results; these are not the author’s self-tests."
          ],
          "artifact_sha256": "d8c7f69530875f185efc88319697eff8e062dedd8dbb403cc5b62ed97fda0ef7",
          "status": "provisional AI assessment",
          "rationale": "AI review disposition: verification only",
          "served_artifact_sha256": "d2346bef1508fb2026e010ceaf60b66382532c82d6efb570249404faeb602cd6",
          "review_evidence_id": "essai03-separate-review-08-14"
        }
      },
      "status": "first attempt retained"
    },
    {
      "id": "BPB2-14",
      "self": {
        "initial": {
          "reviewer": "gpt-6-astra · answering agent",
          "method": "Unverified self-assessment before feedback; common assessment/source packaging added afterward",
          "assessed_at": "2026-11-14T18:41:11.178137+00:00",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 3,
            "B": 3
          },
          "reasons": {
            "C": "Derives 24 standard applicable cases and 17 documented fast cases; preserves the seven-case plus 100-retry gate. Computes complete parallel rollback with rounded replay and 200 in-flight, including standard 2350 boundary. Preserves pending refunds and authority limits.",
            "D": "Recommends feasible fast preparation at existing 2800 cap, 43-minute rollback, explicit 1945/2000/2130 stop gates, bounded standard contingency, costs and reversal conditions.",
            "E": "Material claims cite original retained source records. Recorded, forecast and hypothetical states stay separate, with immutable 18:00 assessment and no invented approvals/results/probabilities.",
            "F": "Includes recommendation, evidence applicability, transactions and aggregates, rollback stage timeline, queue-range analysis, cost comparison, staffing, authority, deadlines, reset and source access.",
            "M": "Leads with named audience and recommendation, then practical conditions and inspectable detail. Supporting content is long and may require scrolling.",
            "B": "21 local JSDOM assertions passed without runtime errors; controls update dependent results and preserve sources. Actual browser rendering and narrow-screen layout remain unverified."
          },
          "material_failures": [],
          "rationale": "Self-assessment only.",
          "artifact_sha256": "96880b2e2d37a09cc06bdd53ff0d95b53c0b13197e3ea0bc1e0f8b2f534b76c9",
          "verification_gaps": []
        },
        "current": {
          "reviewer": "gpt-6-astra · answering agent",
          "method": "Unverified self-assessment before feedback; common assessment/source packaging added afterward; current case analysis unchanged",
          "assessed_at": "2026-11-14T18:41:11.178137+00:00",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 3,
            "B": 3
          },
          "reasons": {
            "C": "Derives 24 standard applicable cases and 17 documented fast cases; preserves the seven-case plus 100-retry gate. Computes complete parallel rollback with rounded replay and 200 in-flight, including standard 2350 boundary. Preserves pending refunds and authority limits.",
            "D": "Recommends feasible fast preparation at existing 2800 cap, 43-minute rollback, explicit 1945/2000/2130 stop gates, bounded standard contingency, costs and reversal conditions.",
            "E": "Material claims cite original retained source records. Recorded, forecast and hypothetical states stay separate, with immutable 18:00 assessment and no invented approvals/results/probabilities.",
            "F": "Includes recommendation, evidence applicability, transactions and aggregates, rollback stage timeline, queue-range analysis, cost comparison, staffing, authority, deadlines, reset and source access.",
            "M": "Leads with named audience and recommendation, then practical conditions and inspectable detail. Supporting content is long and may require scrolling.",
            "B": "21 local JSDOM assertions passed without runtime errors; controls update dependent results and preserve sources. Actual browser rendering and narrow-screen layout remain unverified."
          },
          "material_failures": [],
          "rationale": "Self-assessment only.",
          "artifact_sha256": "96880b2e2d37a09cc06bdd53ff0d95b53c0b13197e3ea0bc1e0f8b2f534b76c9",
          "verification_gaps": []
        }
      },
      "checked": {
        "initial": {
          "reviewer": "gpt-6-astra · separate assessment agent",
          "method": "Separate source/logic review, independent checks and coordinator ordinary-browser observations; same model family, not a human or Claude assessment",
          "assessed_at": "2026-11-14T18:49:43.095990Z",
          "case_id": "BPB2-14",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 4,
            "B": 4
          },
          "reasons": {
            "C": "Recomputed standard24 applicable critical cases and fast17 plus seven reruns/100retries, transaction settlement and staffing. Independent timing tests over210 worker/cap/count combinations match the page: standard2350→45min,2351→46; fast2800→43. In-flight200 and parallel stages are correctly included.",
            "D": "Conditional £6,000 fast2800 route is feasible with19:00 tests,19:45 acceptance,20:00 decision and21:30 stop. Explains tighter standard2350 alternative,£42,000 deferral/£48,000 afterfastfailure and unpriced cash-only disruption; explicit reversal conditions avoid invented expected-value assumptions.",
            "E": "Accepts Q101 sale/basic refund and Q114 replay by relevant fingerprints, retains pending RF88 and genuine settled/pending classifications, and preserves frozen18:00 assessment. Future test/approval controls are expressly hypothetical; cap changes invalidate dependent assumptions.",
            "F": "Configuration evidence, transaction treatment, full rollback stages, enforced versus observed queue, staffing, preparation authority, commercial consequences and time-bounded release/stop route form a complete decision record.",
            "M": "Recommendation, frozen assessment and staged decision deadline sequence are distinct. Queue bound versus actual count and captured/settled/future net amounts sit beside explanations; comparison makes standard technical support versus disruption readily inspectable.",
            "B": "Coordinator ordinary-browser checked2351standard, highqueuehold, changed-cap clearing, failedfast persistence, invalid3201 andreset. Independent210 state timing checks plus accepted/failed/reset/invalid paths confirmed coherent outputs, honest stops and restored initial assumptions."
          },
          "material_failures": [],
          "findings": [
            {
              "severity": "observation",
              "description": "Candidate key gives the fast arithmetic ceiling3200; answer distinguishes that from supplied authority to reduce the current2800cap. Keeping2800 and requiring fresh authority for an increase is a defensible conservative reading, not a calculation failure.",
              "evidence": "Source pack, answer and candidate key compared after independent interpretation; see independent-interpretation.md and independent checks.",
              "proposed_repair": "No repair required on this point; retain the defensible alternative rather than force key wording or allocation."
            }
          ],
          "verification_gaps": [
            "This is a separate AI assessment within the same model family, not Claude or human professional validation.",
            "Native browser evidence is the coordinator’s bounded ordinary-browser observations; the assessor independently ran JSDOM checks, which do not establish comprehensive rendering, keyboard or assistive-technology coverage."
          ],
          "disposition": "verification only",
          "independent_checks": [
            "210 combinations: worker1/2;15 selectedcaps including threshold neighbours;7 queuecounts. All bound/observed timings and hard stops match independent formula28+max(ceil((queue+200)/rate),12).",
            "RehearsalH1800alreadytotal→40min; standardceiling2350queued; fastceiling3200queued. Capabove2800 separately held for unprovided upward authority.",
            "Captured320,settled260,pendingincoming60,pendingrefund80,futureconditionalnet240;154/160=96.25%;deferral42000/fastthen48000.",
            "Hypothetical all-gates state, cap-change invalidation, failedfast retained onstandardswitch, invalidblank/negative/fractional/3201 and reset exercised.",
            "Reviewer DOM outcomes: 219 passed checks, 0 failed assertions. See saved checking code and results; these are not the author’s self-tests."
          ],
          "artifact_sha256": "96880b2e2d37a09cc06bdd53ff0d95b53c0b13197e3ea0bc1e0f8b2f534b76c9",
          "status": "provisional AI assessment",
          "rationale": "AI review disposition: verification only",
          "served_artifact_sha256": "e20eb35910c4f9d35b8c32b405784e4133f78a59a82e973c9cbb6ea229c1480d",
          "review_evidence_id": "essai03-separate-review-08-14"
        },
        "current": {
          "reviewer": "gpt-6-astra · separate assessment agent",
          "method": "Separate source/logic review, independent checks and coordinator ordinary-browser observations; same model family, not a human or Claude assessment; current case analysis is unchanged from the retained first answer",
          "assessed_at": "2026-11-14T18:49:43.095990Z",
          "case_id": "BPB2-14",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 4,
            "B": 4
          },
          "reasons": {
            "C": "Recomputed standard24 applicable critical cases and fast17 plus seven reruns/100retries, transaction settlement and staffing. Independent timing tests over210 worker/cap/count combinations match the page: standard2350→45min,2351→46; fast2800→43. In-flight200 and parallel stages are correctly included.",
            "D": "Conditional £6,000 fast2800 route is feasible with19:00 tests,19:45 acceptance,20:00 decision and21:30 stop. Explains tighter standard2350 alternative,£42,000 deferral/£48,000 afterfastfailure and unpriced cash-only disruption; explicit reversal conditions avoid invented expected-value assumptions.",
            "E": "Accepts Q101 sale/basic refund and Q114 replay by relevant fingerprints, retains pending RF88 and genuine settled/pending classifications, and preserves frozen18:00 assessment. Future test/approval controls are expressly hypothetical; cap changes invalidate dependent assumptions.",
            "F": "Configuration evidence, transaction treatment, full rollback stages, enforced versus observed queue, staffing, preparation authority, commercial consequences and time-bounded release/stop route form a complete decision record.",
            "M": "Recommendation, frozen assessment and staged decision deadline sequence are distinct. Queue bound versus actual count and captured/settled/future net amounts sit beside explanations; comparison makes standard technical support versus disruption readily inspectable.",
            "B": "Coordinator ordinary-browser checked2351standard, highqueuehold, changed-cap clearing, failedfast persistence, invalid3201 andreset. Independent210 state timing checks plus accepted/failed/reset/invalid paths confirmed coherent outputs, honest stops and restored initial assumptions."
          },
          "material_failures": [],
          "findings": [
            {
              "severity": "observation",
              "description": "Candidate key gives the fast arithmetic ceiling3200; answer distinguishes that from supplied authority to reduce the current2800cap. Keeping2800 and requiring fresh authority for an increase is a defensible conservative reading, not a calculation failure.",
              "evidence": "Source pack, answer and candidate key compared after independent interpretation; see independent-interpretation.md and independent checks.",
              "proposed_repair": "No repair required on this point; retain the defensible alternative rather than force key wording or allocation."
            }
          ],
          "verification_gaps": [
            "This is a separate AI assessment within the same model family, not Claude or human professional validation.",
            "Native browser evidence is the coordinator’s bounded ordinary-browser observations; the assessor independently ran JSDOM checks, which do not establish comprehensive rendering, keyboard or assistive-technology coverage."
          ],
          "disposition": "verification only",
          "independent_checks": [
            "210 combinations: worker1/2;15 selectedcaps including threshold neighbours;7 queuecounts. All bound/observed timings and hard stops match independent formula28+max(ceil((queue+200)/rate),12).",
            "RehearsalH1800alreadytotal→40min; standardceiling2350queued; fastceiling3200queued. Capabove2800 separately held for unprovided upward authority.",
            "Captured320,settled260,pendingincoming60,pendingrefund80,futureconditionalnet240;154/160=96.25%;deferral42000/fastthen48000.",
            "Hypothetical all-gates state, cap-change invalidation, failedfast retained onstandardswitch, invalidblank/negative/fractional/3201 and reset exercised.",
            "Reviewer DOM outcomes: 219 passed checks, 0 failed assertions. See saved checking code and results; these are not the author’s self-tests."
          ],
          "artifact_sha256": "96880b2e2d37a09cc06bdd53ff0d95b53c0b13197e3ea0bc1e0f8b2f534b76c9",
          "status": "provisional AI assessment",
          "rationale": "AI review disposition: verification only",
          "served_artifact_sha256": "05531018668667a5f9384b87cd6d638c4f9a4a70f1f9940a975680c3e089e9b3",
          "review_evidence_id": "essai03-separate-review-08-14"
        }
      },
      "status": "first attempt retained"
    },
    {
      "id": "BPB2-15",
      "self": {
        "initial": {
          "reviewer": "gpt-6-astra · answering agent",
          "method": "Unverified self-assessment before feedback; common assessment/source packaging added afterward",
          "assessed_at": "2026-11-14T18:51:26Z",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 3,
            "B": 3
          },
          "reasons": {
            "C": "Uses the family-specific identifier matrix, correct batch/yield arithmetic, finite named cover, training-date and rest constraints, severable milestones and five-weekday payment calculations. No unsupported certificate or future pass is claimed.",
            "D": "Recommends a concrete early-A route with named cover, an explicit no-night 10 Aug decision, £3,200 net comparative benefit, cash consequences, owners and fallback gates.",
            "E": "Fixed 4 Aug facts remain visible; future outcomes and approvals are scenario assumptions. Source IDs link to embedded supporting records. Unknown later bookings and actual production calendar remain unknown.",
            "F": "Covers installed, temporary, provisional A/B and final states; evidence reuse, dated plan, packing limits, discount, cash, warranties, minor closure and test/training failures. 9 Aug failure moves the next opportunities and payment gates while retaining valid evidence.",
            "M": "Professional decision-record structure, recommendation first, comparison table/benefit bars, source details and visible working. It is substantial and long; real browser visual readability remains unverified.",
            "B": "40 local JSDOM assertions passed and 256 input combinations rendered without script errors or undefined results. Reset, scenario dependencies and source expansion work in DOM. No real rendered-browser or responsive-layout check was available."
          },
          "material_failures": [],
          "rationale": "Self-assessment only.",
          "artifact_sha256": "4e67b77411b2dc64133d146d17c82cd8952c527dd5baf181d007dec4ef37d245",
          "verification_gaps": []
        },
        "current": {
          "reviewer": "gpt-6-astra · answering agent",
          "method": "Unverified self-assessment before feedback; common assessment/source packaging added afterward; current case analysis unchanged",
          "assessed_at": "2026-11-14T18:51:26Z",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 3,
            "B": 3
          },
          "reasons": {
            "C": "Uses the family-specific identifier matrix, correct batch/yield arithmetic, finite named cover, training-date and rest constraints, severable milestones and five-weekday payment calculations. No unsupported certificate or future pass is claimed.",
            "D": "Recommends a concrete early-A route with named cover, an explicit no-night 10 Aug decision, £3,200 net comparative benefit, cash consequences, owners and fallback gates.",
            "E": "Fixed 4 Aug facts remain visible; future outcomes and approvals are scenario assumptions. Source IDs link to embedded supporting records. Unknown later bookings and actual production calendar remain unknown.",
            "F": "Covers installed, temporary, provisional A/B and final states; evidence reuse, dated plan, packing limits, discount, cash, warranties, minor closure and test/training failures. 9 Aug failure moves the next opportunities and payment gates while retaining valid evidence.",
            "M": "Professional decision-record structure, recommendation first, comparison table/benefit bars, source details and visible working. It is substantial and long; real browser visual readability remains unverified.",
            "B": "40 local JSDOM assertions passed and 256 input combinations rendered without script errors or undefined results. Reset, scenario dependencies and source expansion work in DOM. No real rendered-browser or responsive-layout check was available."
          },
          "material_failures": [],
          "rationale": "Self-assessment only.",
          "artifact_sha256": "4e67b77411b2dc64133d146d17c82cd8952c527dd5baf181d007dec4ef37d245",
          "verification_gaps": []
        }
      },
      "checked": {
        "initial": {
          "case_id": "BPB2-15",
          "reviewer": "gpt-6-astra · separate assessment agent",
          "method": "Separate source/logic review, independent checks and coordinator ordinary-browser observations; same model family, not a human or Claude assessment",
          "assessed_at": "2026-11-14T18:58:26.756115+00:00",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 4,
            "B": 4
          },
          "reasons": {
            "C": "Independently reproduced A yields99.25/99/98.75%, transferable M7 throughput/B2 barcode and required T2/interface renewal. Early/temporary net benefits1200/3000 plus5000 discount give3200 advantage. Staffing prevents the three-post/two-person10-Aug collision, and uses at most five B cover nights. Certificate/invoice and bounded failure scenarios are coherent across256 control combinations.",
            "D": "Makes a defensible early-A recommendation, prices the alternative at3200, names Maeve/Tomas/Hana and exposes low cash25000, earlier warranty and valid-certificate conditions. Provides temporary fallback when cover is withheld and specific holds after failed tests/training.",
            "E": "Separates signed old evidence, retained identifier-specific applicability, forecast success, proposed cover and actual approvals. No future certification, food-safety authority or training result is invented; finite cover availability is retained after delay.",
            "F": "Installed, temporary, provisional A/B and final states are distinct; interface, legacy output, approvals, rotas, payments, warranty and actual-production final clock form a coherent handover record.",
            "M": "Decision and cash tradeoff lead the page; named sections expose evidence families, rota conflicts, payment workings and original source position. Conditional scenario labelling prevents hypothetical success being mistaken for recorded acceptance.",
            "B": "Coordinator ordinary-browser evidence covers10-Aug infeasible night,9/16-Aug failed tests, withheld-cover fallback and reload. Independent JSDOM checks cover all256 selections, payment/date propagation, reset and local anchors. Required9-Aug failure interaction preserves old evidence and revises16/17-Aug opportunities without extending cover."
          },
          "material_failures": [],
          "findings": [],
          "verification_gaps": [
            "This is a separate same-family AI assessment, not practitioner checking, Claude validation or real-world authority.",
            "Native browser observations were supplied by the coordinator; the reviewer independently ran JSDOM and source-derived checks, not a separate native-browser session.",
            "No claim of exhaustive browser-engine, keyboard-accessibility or responsive-layout coverage."
          ],
          "disposition": "verification only",
          "independent_checks": [
            "case15-results.json: four independent check groups pass, including256 scenario combinations.",
            "Output/discount arithmetic: early1200 net production+5000 discount vs temporary3000 =>3200 advantage; total discounted contract895000.",
            "Payment dates excluding certificate:4-Aug→11-Aug ordinarily,10-Aug→17-Aug,17-Aug→24-Aug; final30 weekday illustration21-Sep/28-Sep is conditional on actual production.",
            "Named cover: A5/6/9, B11/12/13/16/17; no10-Aug integrated night; night-to-next-day rest6h<11h."
          ],
          "artifact_sha256": "4e67b77411b2dc64133d146d17c82cd8952c527dd5baf181d007dec4ef37d245",
          "status": "provisional AI assessment",
          "rationale": "AI review disposition: verification only",
          "served_artifact_sha256": "bf53de0b0ace6fe306a1f7ec49f4af3d88a9f3cd6be15c31dfe768d43bc339c7",
          "review_evidence_id": "essai03-separate-review-15-20"
        },
        "current": {
          "case_id": "BPB2-15",
          "reviewer": "gpt-6-astra · separate assessment agent",
          "method": "Separate source/logic review, independent checks and coordinator ordinary-browser observations; same model family, not a human or Claude assessment; current case analysis is unchanged from the retained first answer",
          "assessed_at": "2026-11-14T18:58:26.756115+00:00",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 4,
            "B": 4
          },
          "reasons": {
            "C": "Independently reproduced A yields99.25/99/98.75%, transferable M7 throughput/B2 barcode and required T2/interface renewal. Early/temporary net benefits1200/3000 plus5000 discount give3200 advantage. Staffing prevents the three-post/two-person10-Aug collision, and uses at most five B cover nights. Certificate/invoice and bounded failure scenarios are coherent across256 control combinations.",
            "D": "Makes a defensible early-A recommendation, prices the alternative at3200, names Maeve/Tomas/Hana and exposes low cash25000, earlier warranty and valid-certificate conditions. Provides temporary fallback when cover is withheld and specific holds after failed tests/training.",
            "E": "Separates signed old evidence, retained identifier-specific applicability, forecast success, proposed cover and actual approvals. No future certification, food-safety authority or training result is invented; finite cover availability is retained after delay.",
            "F": "Installed, temporary, provisional A/B and final states are distinct; interface, legacy output, approvals, rotas, payments, warranty and actual-production final clock form a coherent handover record.",
            "M": "Decision and cash tradeoff lead the page; named sections expose evidence families, rota conflicts, payment workings and original source position. Conditional scenario labelling prevents hypothetical success being mistaken for recorded acceptance.",
            "B": "Coordinator ordinary-browser evidence covers10-Aug infeasible night,9/16-Aug failed tests, withheld-cover fallback and reload. Independent JSDOM checks cover all256 selections, payment/date propagation, reset and local anchors. Required9-Aug failure interaction preserves old evidence and revises16/17-Aug opportunities without extending cover."
          },
          "material_failures": [],
          "findings": [],
          "verification_gaps": [
            "This is a separate same-family AI assessment, not practitioner checking, Claude validation or real-world authority.",
            "Native browser observations were supplied by the coordinator; the reviewer independently ran JSDOM and source-derived checks, not a separate native-browser session.",
            "No claim of exhaustive browser-engine, keyboard-accessibility or responsive-layout coverage."
          ],
          "disposition": "verification only",
          "independent_checks": [
            "case15-results.json: four independent check groups pass, including256 scenario combinations.",
            "Output/discount arithmetic: early1200 net production+5000 discount vs temporary3000 =>3200 advantage; total discounted contract895000.",
            "Payment dates excluding certificate:4-Aug→11-Aug ordinarily,10-Aug→17-Aug,17-Aug→24-Aug; final30 weekday illustration21-Sep/28-Sep is conditional on actual production.",
            "Named cover: A5/6/9, B11/12/13/16/17; no10-Aug integrated night; night-to-next-day rest6h<11h."
          ],
          "artifact_sha256": "4e67b77411b2dc64133d146d17c82cd8952c527dd5baf181d007dec4ef37d245",
          "status": "provisional AI assessment",
          "rationale": "AI review disposition: verification only",
          "served_artifact_sha256": "fb8d68053141bfa6596f31595f2af46ba1acaef4971bf2635b02680f4ebd9619",
          "review_evidence_id": "essai03-separate-review-15-20"
        }
      },
      "status": "first attempt retained"
    },
    {
      "id": "BPB2-16",
      "self": {
        "initial": {
          "reviewer": "gpt-6-astra · answering agent",
          "method": "Unverified self-assessment before feedback; common assessment/source packaging added afterward",
          "assessed_at": "2026-11-14T17:53:53Z",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 3,
            "B": 3
          },
          "reasons": {
            "C": "Exact-fraction arithmetic check confirms weather-only heating normalisation, weighted unscaled C-B correction, T-C counterfactual increment, all eight T-D bounds, fixed 625,000 denominator and -21,750/+61,450 extrema.",
            "D": "Leads with a committee-ready bounded claim and explicit defer-payment recommendation; prioritises the fully costed £3,000 programme with owners, June 18 return and quarterly review.",
            "E": "Source-linked records remain available; missing meter observations stay unknown; scenarios are labelled hypothetical with no distributions, confidence or causal certainty; actual bills, bill equivalents and comfort evidence are separate.",
            "F": "Includes headline audit, normalised P0/P1/P2 bridge, cohort/portfolio bounds, linked scenario controls, billing decomposition, comfort coverage, monitoring/payment plan, reset and original evidence recovery.",
            "M": "Clear recommendation and section navigation, annotated tables and a range chart; the complete calculation and evidence trail makes the page long and has not been visually reviewed in a real browser.",
            "B": "Twelve JSDOM checks passed for initial state, controls, corners, invalid input, reset, reload and evidence links, without script errors. Real-browser rendering, responsive behaviour and keyboard journeys remain unverified."
          },
          "material_failures": [],
          "rationale": "Self-assessment only.",
          "artifact_sha256": "65b8303be9f43b309148ffe3e7a4e4534311ff482d16f09a64e94a87d4fda83e",
          "verification_gaps": []
        },
        "current": {
          "reviewer": "gpt-6-astra · answering agent",
          "method": "Unverified self-assessment before feedback; common assessment/source packaging added afterward; current case analysis unchanged",
          "assessed_at": "2026-11-14T17:53:53Z",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 3,
            "B": 3
          },
          "reasons": {
            "C": "Exact-fraction arithmetic check confirms weather-only heating normalisation, weighted unscaled C-B correction, T-C counterfactual increment, all eight T-D bounds, fixed 625,000 denominator and -21,750/+61,450 extrema.",
            "D": "Leads with a committee-ready bounded claim and explicit defer-payment recommendation; prioritises the fully costed £3,000 programme with owners, June 18 return and quarterly review.",
            "E": "Source-linked records remain available; missing meter observations stay unknown; scenarios are labelled hypothetical with no distributions, confidence or causal certainty; actual bills, bill equivalents and comfort evidence are separate.",
            "F": "Includes headline audit, normalised P0/P1/P2 bridge, cohort/portfolio bounds, linked scenario controls, billing decomposition, comfort coverage, monitoring/payment plan, reset and original evidence recovery.",
            "M": "Clear recommendation and section navigation, annotated tables and a range chart; the complete calculation and evidence trail makes the page long and has not been visually reviewed in a real browser.",
            "B": "Twelve JSDOM checks passed for initial state, controls, corners, invalid input, reset, reload and evidence links, without script errors. Real-browser rendering, responsive behaviour and keyboard journeys remain unverified."
          },
          "material_failures": [],
          "rationale": "Self-assessment only.",
          "artifact_sha256": "65b8303be9f43b309148ffe3e7a4e4534311ff482d16f09a64e94a87d4fda83e",
          "verification_gaps": []
        }
      },
      "checked": {
        "initial": {
          "case_id": "BPB2-16",
          "reviewer": "gpt-6-astra · separate assessment agent",
          "method": "Separate source/logic review, independent checks and coordinator ordinary-browser observations; same model family, not a human or Claude assessment",
          "assessed_at": "2026-11-14T18:58:26.756115+00:00",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 4,
            "B": 4
          },
          "reasons": {
            "C": "Independently recomputed all eight uncertainty corners and an interior setting: fixed denominator625000; portfolio -21750 to61450kWh, -3.48% to9.832%; avoided bills -2392.50 to6759.50; carbon -4.002 to11.3068t. Same40 observed bills40270 to48942, increase8672. Pre-trend and cohort bridges reconcile.",
            "D": "Clear payment hold, all3000 of specified monitoring, greatest-uncertainty priority, named ownership and18-Jun return. Explains that narrowing current bounds cannot itself reach20%, and preserves the committee decision rather than inventing automatic clawback.",
            "E": "Audits the48/40 population mismatch, known boiler contamination, deterministic bounds, unknown meters and voluntary comfort response. Retains source claims, baseline, uncertainty and committee authority throughout.",
            "F": "Claim audit, cohort bridge, target bounds, actual/current-price bills, carbon, pre-trends, comfort, monitoring capacity and owned payment decision are coherently connected.",
            "M": "Decision-first claim with fixed original range; labelled scenario controls, cohort detail, billing decomposition and source disclosures keep denominators, units and uncertainty close to claims.",
            "B": "Coordinator observed upper-bound recalculation, invalid meter rejection, reset and actual-bill separation. Independent JSDOM reproduced all corner/interior energy-money-carbon propagation, invalid/blank inputs, reset and all rendered anchors. Scope is these inputs and reported ordinary-browser journeys."
          },
          "material_failures": [],
          "findings": [],
          "verification_gaps": [
            "This is a separate same-family AI assessment, not practitioner checking, Claude validation or real-world authority.",
            "Native browser observations were supplied by the coordinator; the reviewer independently ran JSDOM and source-derived checks, not a separate native-browser session.",
            "No claim of exhaustive browser-engine, keyboard-accessibility or responsive-layout coverage."
          ],
          "disposition": "verification only",
          "independent_checks": [
            "independent-results.json: eight bound corners and one interior setting match source-derived energy, percentage, bill and carbon calculations.",
            "Same40 bills40270→48942; P0 normalisations independently reconstructed.",
            "Invalid empty/above-range/negative boiler entries retain labelled last-valid results; reset restores -21750."
          ],
          "artifact_sha256": "65b8303be9f43b309148ffe3e7a4e4534311ff482d16f09a64e94a87d4fda83e",
          "status": "provisional AI assessment",
          "rationale": "AI review disposition: verification only",
          "served_artifact_sha256": "b62f1e066bf2e51175929c1643739e9eaa7c506dfcd7d4366ae39a6ebca9541f",
          "review_evidence_id": "essai03-separate-review-15-20"
        },
        "current": {
          "case_id": "BPB2-16",
          "reviewer": "gpt-6-astra · separate assessment agent",
          "method": "Separate source/logic review, independent checks and coordinator ordinary-browser observations; same model family, not a human or Claude assessment; current case analysis is unchanged from the retained first answer",
          "assessed_at": "2026-11-14T18:58:26.756115+00:00",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 4,
            "B": 4
          },
          "reasons": {
            "C": "Independently recomputed all eight uncertainty corners and an interior setting: fixed denominator625000; portfolio -21750 to61450kWh, -3.48% to9.832%; avoided bills -2392.50 to6759.50; carbon -4.002 to11.3068t. Same40 observed bills40270 to48942, increase8672. Pre-trend and cohort bridges reconcile.",
            "D": "Clear payment hold, all3000 of specified monitoring, greatest-uncertainty priority, named ownership and18-Jun return. Explains that narrowing current bounds cannot itself reach20%, and preserves the committee decision rather than inventing automatic clawback.",
            "E": "Audits the48/40 population mismatch, known boiler contamination, deterministic bounds, unknown meters and voluntary comfort response. Retains source claims, baseline, uncertainty and committee authority throughout.",
            "F": "Claim audit, cohort bridge, target bounds, actual/current-price bills, carbon, pre-trends, comfort, monitoring capacity and owned payment decision are coherently connected.",
            "M": "Decision-first claim with fixed original range; labelled scenario controls, cohort detail, billing decomposition and source disclosures keep denominators, units and uncertainty close to claims.",
            "B": "Coordinator observed upper-bound recalculation, invalid meter rejection, reset and actual-bill separation. Independent JSDOM reproduced all corner/interior energy-money-carbon propagation, invalid/blank inputs, reset and all rendered anchors. Scope is these inputs and reported ordinary-browser journeys."
          },
          "material_failures": [],
          "findings": [],
          "verification_gaps": [
            "This is a separate same-family AI assessment, not practitioner checking, Claude validation or real-world authority.",
            "Native browser observations were supplied by the coordinator; the reviewer independently ran JSDOM and source-derived checks, not a separate native-browser session.",
            "No claim of exhaustive browser-engine, keyboard-accessibility or responsive-layout coverage."
          ],
          "disposition": "verification only",
          "independent_checks": [
            "independent-results.json: eight bound corners and one interior setting match source-derived energy, percentage, bill and carbon calculations.",
            "Same40 bills40270→48942; P0 normalisations independently reconstructed.",
            "Invalid empty/above-range/negative boiler entries retain labelled last-valid results; reset restores -21750."
          ],
          "artifact_sha256": "65b8303be9f43b309148ffe3e7a4e4534311ff482d16f09a64e94a87d4fda83e",
          "status": "provisional AI assessment",
          "rationale": "AI review disposition: verification only",
          "served_artifact_sha256": "f9ca4acb73b4fbffa122e397db0ab5ac1d3327938edb7428845096115047aa0b",
          "review_evidence_id": "essai03-separate-review-15-20"
        }
      },
      "status": "first attempt retained"
    },
    {
      "id": "BPB2-17",
      "self": {
        "initial": {
          "reviewer": "gpt-6-astra · answering agent",
          "method": "Unverified self-assessment before feedback; common assessment/source packaging added afterward",
          "assessed_at": "2026-11-14T18:12:30.987730Z",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 3,
            "B": 3
          },
          "reasons": {
            "C": "Six daily witnesses satisfy stated effort, dependencies, calendars, observation timing, non-anticipation, windows and gross authority in local mechanical checks.",
            "D": "Recommends S+K and D26 with explicit tradeoff against executable £32k earliest policy; both probability views and the 30.769% economic switch are visible.",
            "E": "Source IDs and original evidence retained; facts, forecasts, proposals and gates separated; external value never funds project cash; bounded future assumptions explicit.",
            "F": "Includes purchase baskets, all state branches, daily resources, funding and protected care, observations, failed alternatives, readiness reconciliation, owners and reversal points.",
            "M": "Outcome-led brief, accessible labelled controls, comparison tables and cash strip are clear; the complete daily witnesses and evidence make the page long.",
            "B": "All 12 policy/state/view combinations, reset, evidence anchors and dependent results pass JSDOM tests. Real browser visual rendering was not available and has not been claimed."
          },
          "material_failures": [],
          "rationale": "Self-assessment only.",
          "artifact_sha256": "1c0bb4b447652944865051d9421418ff0007852a241e6c9c557c1d0d60962e59",
          "verification_gaps": []
        },
        "current": {
          "reviewer": "gpt-6-astra · answering agent",
          "method": "Unverified self-assessment before feedback; common assessment/source packaging added afterward; current case analysis unchanged",
          "assessed_at": "2026-11-14T18:12:30.987730Z",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 3,
            "B": 3
          },
          "reasons": {
            "C": "Six daily witnesses satisfy stated effort, dependencies, calendars, observation timing, non-anticipation, windows and gross authority in local mechanical checks.",
            "D": "Recommends S+K and D26 with explicit tradeoff against executable £32k earliest policy; both probability views and the 30.769% economic switch are visible.",
            "E": "Source IDs and original evidence retained; facts, forecasts, proposals and gates separated; external value never funds project cash; bounded future assumptions explicit.",
            "F": "Includes purchase baskets, all state branches, daily resources, funding and protected care, observations, failed alternatives, readiness reconciliation, owners and reversal points.",
            "M": "Outcome-led brief, accessible labelled controls, comparison tables and cash strip are clear; the complete daily witnesses and evidence make the page long.",
            "B": "All 12 policy/state/view combinations, reset, evidence anchors and dependent results pass JSDOM tests. Real browser visual rendering was not available and has not been claimed."
          },
          "material_failures": [],
          "rationale": "Self-assessment only.",
          "artifact_sha256": "1c0bb4b447652944865051d9421418ff0007852a241e6c9c557c1d0d60962e59",
          "verification_gaps": []
        }
      },
      "checked": {
        "initial": {
          "case_id": "BPB2-17",
          "reviewer": "gpt-6-astra · separate assessment agent",
          "method": "Separate source/logic review, independent checks and coordinator ordinary-browser observations; same model family, not a human or Claude assessment",
          "assessed_at": "2026-11-14T18:58:26.756115+00:00",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 4,
            "B": 4
          },
          "reasons": {
            "C": "Independently validated all six resource witnesses: exact effort/roles, F calendars, dependencies, conversion start, UAT restart, no double-booking and nonanticipation. Gross104/144/144 early and128/128/128 lower preserve12 hypercare. Delivery economic115.4 versus128 and finance128.4 versus128, switch16/52, reconcile.",
            "D": "Recommends lower exposure on uniform date and22 funding margin, explicitly giving up12.6 expected economic advantage under delivery while winning only0.4 under finance. Earliest is retained as a defensible alternative with funded fallbacks, decision criteria, owners and reversal points.",
            "E": "Readiness claims reconciled to65% accepted records and unknown core-ticket denominator; probability views kept separate. Schedules are bookings, not evidence passes. K activation, purchases, protected hypercare and outside contribution retain their authority/information boundaries.",
            "F": "Upfront baskets, observation tree, six daily witnesses, expected views, gross state gates, failed substitutes, operational burden and final finance/test/rollback/operations gates form an executable bounded policy comparison.",
            "M": "Priority decision and why precede controls. Full state matrix, expected comparison and daily witness have clear roles; economic comparison is visibly separated from gross funding and protected care.",
            "B": "Coordinator ordinary-browser checks covered early/J3, lower/J3, finance-view change, UAT restart and reset/reload. Independent JSDOM covered all12 policy/state/view selections and verified schedule/funding propagation and anchors. Arbitrary-purchase validation is correctly not offered or required."
          },
          "material_failures": [],
          "findings": [
            {
              "severity": "observation",
              "description": "The submitted schedule differs from the candidate key but is a valid alternative witness. Clean early A endsD7/B endsD8; lower clean A endsD7/B endsD9. The candidate reverses those endpoint allocations.",
              "evidence": "Independent resource, dependency and pre-observation checks in independent-results.json pass all six witnesses.17-C–F permit gaps and post-D6 branching.",
              "proposed_repair": "None. Accept the demonstrated equivalent resource allocation; do not require the candidate exact dates."
            }
          ],
          "verification_gaps": [
            "This is a separate same-family AI assessment, not practitioner checking, Claude validation or real-world authority.",
            "Native browser observations were supplied by the coordinator; the reviewer independently ran JSDOM and source-derived checks, not a separate native-browser session.",
            "No claim of exhaustive browser-engine, keyboard-accessibility or responsive-layout coverage."
          ],
          "disposition": "verification only",
          "independent_checks": [
            "independent-results.json: independent six-witness resource/calendar/dependency checker passes.",
            "Identical decisions throughD6 in all states and through the second actual U day forJ1/J3; verified by event comparison.",
            "Expected costs and threshold independently recomputed; all12 required selector combinations update rendered output."
          ],
          "artifact_sha256": "1c0bb4b447652944865051d9421418ff0007852a241e6c9c557c1d0d60962e59",
          "status": "provisional AI assessment",
          "rationale": "AI review disposition: verification only",
          "served_artifact_sha256": "c8a715094f426eb9f55286f6e766b32565144e61e2a7882729086f09862a1ef4",
          "review_evidence_id": "essai03-separate-review-15-20"
        },
        "current": {
          "case_id": "BPB2-17",
          "reviewer": "gpt-6-astra · separate assessment agent",
          "method": "Separate source/logic review, independent checks and coordinator ordinary-browser observations; same model family, not a human or Claude assessment; current case analysis is unchanged from the retained first answer",
          "assessed_at": "2026-11-14T18:58:26.756115+00:00",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 4,
            "B": 4
          },
          "reasons": {
            "C": "Independently validated all six resource witnesses: exact effort/roles, F calendars, dependencies, conversion start, UAT restart, no double-booking and nonanticipation. Gross104/144/144 early and128/128/128 lower preserve12 hypercare. Delivery economic115.4 versus128 and finance128.4 versus128, switch16/52, reconcile.",
            "D": "Recommends lower exposure on uniform date and22 funding margin, explicitly giving up12.6 expected economic advantage under delivery while winning only0.4 under finance. Earliest is retained as a defensible alternative with funded fallbacks, decision criteria, owners and reversal points.",
            "E": "Readiness claims reconciled to65% accepted records and unknown core-ticket denominator; probability views kept separate. Schedules are bookings, not evidence passes. K activation, purchases, protected hypercare and outside contribution retain their authority/information boundaries.",
            "F": "Upfront baskets, observation tree, six daily witnesses, expected views, gross state gates, failed substitutes, operational burden and final finance/test/rollback/operations gates form an executable bounded policy comparison.",
            "M": "Priority decision and why precede controls. Full state matrix, expected comparison and daily witness have clear roles; economic comparison is visibly separated from gross funding and protected care.",
            "B": "Coordinator ordinary-browser checks covered early/J3, lower/J3, finance-view change, UAT restart and reset/reload. Independent JSDOM covered all12 policy/state/view selections and verified schedule/funding propagation and anchors. Arbitrary-purchase validation is correctly not offered or required."
          },
          "material_failures": [],
          "findings": [
            {
              "severity": "observation",
              "description": "The submitted schedule differs from the candidate key but is a valid alternative witness. Clean early A endsD7/B endsD8; lower clean A endsD7/B endsD9. The candidate reverses those endpoint allocations.",
              "evidence": "Independent resource, dependency and pre-observation checks in independent-results.json pass all six witnesses.17-C–F permit gaps and post-D6 branching.",
              "proposed_repair": "None. Accept the demonstrated equivalent resource allocation; do not require the candidate exact dates."
            }
          ],
          "verification_gaps": [
            "This is a separate same-family AI assessment, not practitioner checking, Claude validation or real-world authority.",
            "Native browser observations were supplied by the coordinator; the reviewer independently ran JSDOM and source-derived checks, not a separate native-browser session.",
            "No claim of exhaustive browser-engine, keyboard-accessibility or responsive-layout coverage."
          ],
          "disposition": "verification only",
          "independent_checks": [
            "independent-results.json: independent six-witness resource/calendar/dependency checker passes.",
            "Identical decisions throughD6 in all states and through the second actual U day forJ1/J3; verified by event comparison.",
            "Expected costs and threshold independently recomputed; all12 required selector combinations update rendered output."
          ],
          "artifact_sha256": "1c0bb4b447652944865051d9421418ff0007852a241e6c9c557c1d0d60962e59",
          "status": "provisional AI assessment",
          "rationale": "AI review disposition: verification only",
          "served_artifact_sha256": "4c12983b4ca1a1b44ed97fbaa7b58c2468bb08ee31cce703199c4675ecfca625",
          "review_evidence_id": "essai03-separate-review-15-20"
        }
      },
      "status": "first attempt retained"
    },
    {
      "id": "BPB2-18",
      "self": {
        "initial": {
          "reviewer": "gpt-6-astra · answering agent",
          "method": "Unverified self-assessment before feedback; common assessment/source packaging added afterward",
          "assessed_at": "2026-11-14T18:18:49Z",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 3,
            "B": 3
          },
          "reasons": {
            "C": "Separate Python and JavaScript implementations agree on route values, Bayesian posteriors, conditional q1 values and liquidity minima. Tests cover grant order, 20 operating quarters, shifted seasons, ramp, sunk deposit and working-capital recovery.",
            "D": "Leads with the highest expected-PV permitted policy, distinguishes the unconstrained economic optimum, prices immediate-exit and limited-scope alternatives, and exposes cash and attendance thresholds that reverse the route or gate decision.",
            "E": "All ten source records are preserved with nearby references; baseline facts, forecasts, hypothetical edits and proposed approvals are distinguished. Every finding retains all three demand possibilities; no refinancing, perfect signal, upgrade or discretionary abandonment is assumed.",
            "F": "Includes corrected bases, five route/policy comparisons, four findings with likelihoods/posteriors/conditional values/exclusions, all twelve investigation paths, movement-level quarterly trace, terminal/grant timing, sensitivity, required controls and reset.",
            "M": "Summary and decision sequence are clear, with accessible labels, text cash table and responsive shared shell. The detailed policy table requires horizontal scrolling on narrow screens and real-browser visual QA remains outstanding.",
            "B": "Ten calculation/DOM test groups pass with no JSDOM runtime errors, including inputs, reset, invalid edits, scenario selection, reload and links. Real browser layout and interaction have not been independently checked."
          },
          "material_failures": [],
          "rationale": "Self-assessment only.",
          "artifact_sha256": "099891c7789460f865e4889dfa4ad9f6d369acf873042ac6b59e296da6992080",
          "verification_gaps": []
        },
        "current": {
          "reviewer": "gpt-6-astra · answering agent",
          "method": "Unverified self-assessment before feedback; common assessment/source packaging added afterward; current case analysis unchanged",
          "assessed_at": "2026-11-14T18:18:49Z",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 3,
            "B": 3
          },
          "reasons": {
            "C": "Separate Python and JavaScript implementations agree on route values, Bayesian posteriors, conditional q1 values and liquidity minima. Tests cover grant order, 20 operating quarters, shifted seasons, ramp, sunk deposit and working-capital recovery.",
            "D": "Leads with the highest expected-PV permitted policy, distinguishes the unconstrained economic optimum, prices immediate-exit and limited-scope alternatives, and exposes cash and attendance thresholds that reverse the route or gate decision.",
            "E": "All ten source records are preserved with nearby references; baseline facts, forecasts, hypothetical edits and proposed approvals are distinguished. Every finding retains all three demand possibilities; no refinancing, perfect signal, upgrade or discretionary abandonment is assumed.",
            "F": "Includes corrected bases, five route/policy comparisons, four findings with likelihoods/posteriors/conditional values/exclusions, all twelve investigation paths, movement-level quarterly trace, terminal/grant timing, sensitivity, required controls and reset.",
            "M": "Summary and decision sequence are clear, with accessible labels, text cash table and responsive shared shell. The detailed policy table requires horizontal scrolling on narrow screens and real-browser visual QA remains outstanding.",
            "B": "Ten calculation/DOM test groups pass with no JSDOM runtime errors, including inputs, reset, invalid edits, scenario selection, reload and links. Real browser layout and interaction have not been independently checked."
          },
          "material_failures": [],
          "rationale": "Self-assessment only.",
          "artifact_sha256": "099891c7789460f865e4889dfa4ad9f6d369acf873042ac6b59e296da6992080",
          "verification_gaps": []
        }
      },
      "checked": {
        "initial": {
          "case_id": "BPB2-18",
          "reviewer": "gpt-6-astra · separate assessment agent",
          "method": "Separate source/logic review, independent checks and coordinator ordinary-browser observations; same model family, not a human or Claude assessment",
          "assessed_at": "2026-11-14T18:58:26.756115+00:00",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 4,
            "B": 4
          },
          "reasons": {
            "C": "A separately authored cash-event ledger matches every36 action/state/delay path at10 input settings, including quarterly payment-order minima and finding posteriors. Baseline full1080.741405 is excluded by164 cash; permitted investigation743.301651 uses full N+, limited O+, exit negatives, min265. Cash2120/2121 and attendance95% boundaries also reproduce.",
            "D": "Recommends highest-PV permitted policy, exposes full-now economic lead, explicit cautious-exit sacrifice123.301651 and negative false-positive path downside. Shows actionable attendance/cash switches without inventing refinancing or perfect signal information.",
            "E": "Reconciles restricted cash, paid deposit, working capital, grant timing, paid-entry margins, seasonal ramp and terminal recovery. Historic spend and job possibilities are excluded from invented benefits; hypothetical cash edits never create authority.",
            "F": "Immediate full/limited/exit, economic and permitted investigation policies, four findings, conditional values, likelihoods/posteriors, exclusions,12 realised outcomes and within-date cash traces are connected and source-reconciled.",
            "M": "Decision-first metrics and original dated recommendation remain visible; navigable reconciliation, finding matrix, downside and trace provide supporting detail with units and clear cash/value distinction.",
            "B": "Coordinator observed LOq2 cash164, cash2121 full switch, attendance90 exit, negative input handling and reset/reload. Independent JSDOM checks the full model at10 settings and required controls/minimum inspection, invalid retention and reset. No optional editable priors are offered."
          },
          "material_failures": [],
          "findings": [],
          "verification_gaps": [
            "This is a separate same-family AI assessment, not practitioner checking, Claude validation or real-world authority.",
            "Native browser observations were supplied by the coordinator; the reviewer independently ran JSDOM and source-derived checks, not a separate native-browser session.",
            "No claim of exhaustive browser-engine, keyboard-accessibility or responsive-layout coverage."
          ],
          "disposition": "verification only",
          "independent_checks": [
            "independent-results.json: independently authored event ledger checks36 paths at10 attendance/cash settings (360 path comparisons) plus all four finding choices/posteriors.",
            "Full LO immediate minimum164 atq2; paused full normal low minimum265q6; grant and terminal checked after preceding cash movement.",
            "Baseline full1080.741405, limited518.655027, exit620, economic investigation787.588437, permitted743.301651.",
            "At95% attendance exit wins; cash2120 full floor199 excludes and2121 floor200 permits; control invalid-state/reset and minimum trace verified."
          ],
          "artifact_sha256": "099891c7789460f865e4889dfa4ad9f6d369acf873042ac6b59e296da6992080",
          "status": "provisional AI assessment",
          "rationale": "AI review disposition: verification only",
          "served_artifact_sha256": "651a05430891a839c0491a02c3d837833a36630cd874d809abdc2c8de9f4a3e4",
          "review_evidence_id": "essai03-separate-review-15-20"
        },
        "current": {
          "case_id": "BPB2-18",
          "reviewer": "gpt-6-astra · separate assessment agent",
          "method": "Separate source/logic review, independent checks and coordinator ordinary-browser observations; same model family, not a human or Claude assessment; current case analysis is unchanged from the retained first answer",
          "assessed_at": "2026-11-14T18:58:26.756115+00:00",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 4,
            "B": 4
          },
          "reasons": {
            "C": "A separately authored cash-event ledger matches every36 action/state/delay path at10 input settings, including quarterly payment-order minima and finding posteriors. Baseline full1080.741405 is excluded by164 cash; permitted investigation743.301651 uses full N+, limited O+, exit negatives, min265. Cash2120/2121 and attendance95% boundaries also reproduce.",
            "D": "Recommends highest-PV permitted policy, exposes full-now economic lead, explicit cautious-exit sacrifice123.301651 and negative false-positive path downside. Shows actionable attendance/cash switches without inventing refinancing or perfect signal information.",
            "E": "Reconciles restricted cash, paid deposit, working capital, grant timing, paid-entry margins, seasonal ramp and terminal recovery. Historic spend and job possibilities are excluded from invented benefits; hypothetical cash edits never create authority.",
            "F": "Immediate full/limited/exit, economic and permitted investigation policies, four findings, conditional values, likelihoods/posteriors, exclusions,12 realised outcomes and within-date cash traces are connected and source-reconciled.",
            "M": "Decision-first metrics and original dated recommendation remain visible; navigable reconciliation, finding matrix, downside and trace provide supporting detail with units and clear cash/value distinction.",
            "B": "Coordinator observed LOq2 cash164, cash2121 full switch, attendance90 exit, negative input handling and reset/reload. Independent JSDOM checks the full model at10 settings and required controls/minimum inspection, invalid retention and reset. No optional editable priors are offered."
          },
          "material_failures": [],
          "findings": [],
          "verification_gaps": [
            "This is a separate same-family AI assessment, not practitioner checking, Claude validation or real-world authority.",
            "Native browser observations were supplied by the coordinator; the reviewer independently ran JSDOM and source-derived checks, not a separate native-browser session.",
            "No claim of exhaustive browser-engine, keyboard-accessibility or responsive-layout coverage."
          ],
          "disposition": "verification only",
          "independent_checks": [
            "independent-results.json: independently authored event ledger checks36 paths at10 attendance/cash settings (360 path comparisons) plus all four finding choices/posteriors.",
            "Full LO immediate minimum164 atq2; paused full normal low minimum265q6; grant and terminal checked after preceding cash movement.",
            "Baseline full1080.741405, limited518.655027, exit620, economic investigation787.588437, permitted743.301651.",
            "At95% attendance exit wins; cash2120 full floor199 excludes and2121 floor200 permits; control invalid-state/reset and minimum trace verified."
          ],
          "artifact_sha256": "099891c7789460f865e4889dfa4ad9f6d369acf873042ac6b59e296da6992080",
          "status": "provisional AI assessment",
          "rationale": "AI review disposition: verification only",
          "served_artifact_sha256": "c54fa7c432be6bdd51dc597f54073ca255a1a6b3a09200421e682b0ebdb75c5e",
          "review_evidence_id": "essai03-separate-review-15-20"
        }
      },
      "status": "first attempt retained"
    },
    {
      "id": "BPB2-19",
      "self": {
        "initial": {
          "reviewer": "gpt-6-astra · answering agent",
          "method": "Unverified self-assessment before feedback; common assessment/source packaging added afterward",
          "assessed_at": "2026-11-14T18:46:49Z",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 3,
            "B": 3
          },
          "reasons": {
            "C": "Recommended G16+R+I totals £80k and 8 Gia-days. D1 G16 approval gives field D2–D11 and consecutive report D12–D15; D5 approval gives I D6–D9 and R order D6. Live checks enforce £95k cap, £75k authority, field access, evidence cut-off and impossible K interface requirement.",
            "D": "Named proposal, staged £30k exposure and £50k board balance, release instructions for adverse/late/unsupported evidence, contingency when board declines, and a distinct £68k G8+R+I alternative are actionable.",
            "E": "Source disclosures retain all six historical cases and adverse evidence. Audit separates elapsed duration, crew exposure, active crew productivity, off-site boundary and pooled steel purchasing. No invented probability, avoided loss, cash credit or causal certainty.",
            "F": "Selected and rejected lessons each include mechanism, boundary, rival explanation and change trigger. Budget, capacity, approval timing, baseline amendments, owners, sources and ephemeral reset behaviour are supplied.",
            "M": "Recommendation leads, audit uses paired visual bars, and schedule/controls use tables. The memo is deliberately detailed; ordinary-browser visual reading and narrow-screen layout remain unverified.",
            "B": "Fourteen JSDOM checks passed, including package changes, invalid inputs, all ground outcomes, authority, cost/capacity, source disclosures and reset; no JavaScript errors remain. No actual browser rendering or manual keyboard walkthrough was available."
          },
          "material_failures": [],
          "rationale": "Self-assessment only.",
          "artifact_sha256": "7f2f504654d160affa77a72bab6321441de8ed24fb616fa3605d121bb32c6d04",
          "verification_gaps": []
        },
        "current": {
          "reviewer": "gpt-6-astra · answering agent",
          "method": "Unverified self-assessment before feedback; common assessment/source packaging added afterward; current case analysis unchanged",
          "assessed_at": "2026-11-14T18:46:49Z",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 3,
            "B": 3
          },
          "reasons": {
            "C": "Recommended G16+R+I totals £80k and 8 Gia-days. D1 G16 approval gives field D2–D11 and consecutive report D12–D15; D5 approval gives I D6–D9 and R order D6. Live checks enforce £95k cap, £75k authority, field access, evidence cut-off and impossible K interface requirement.",
            "D": "Named proposal, staged £30k exposure and £50k board balance, release instructions for adverse/late/unsupported evidence, contingency when board declines, and a distinct £68k G8+R+I alternative are actionable.",
            "E": "Source disclosures retain all six historical cases and adverse evidence. Audit separates elapsed duration, crew exposure, active crew productivity, off-site boundary and pooled steel purchasing. No invented probability, avoided loss, cash credit or causal certainty.",
            "F": "Selected and rejected lessons each include mechanism, boundary, rival explanation and change trigger. Budget, capacity, approval timing, baseline amendments, owners, sources and ephemeral reset behaviour are supplied.",
            "M": "Recommendation leads, audit uses paired visual bars, and schedule/controls use tables. The memo is deliberately detailed; ordinary-browser visual reading and narrow-screen layout remain unverified.",
            "B": "Fourteen JSDOM checks passed, including package changes, invalid inputs, all ground outcomes, authority, cost/capacity, source disclosures and reset; no JavaScript errors remain. No actual browser rendering or manual keyboard walkthrough was available."
          },
          "material_failures": [],
          "rationale": "Self-assessment only.",
          "artifact_sha256": "7f2f504654d160affa77a72bab6321441de8ed24fb616fa3605d121bb32c6d04",
          "verification_gaps": []
        }
      },
      "checked": {
        "initial": {
          "case_id": "BPB2-19",
          "reviewer": "gpt-6-astra · separate assessment agent",
          "method": "Separate source/logic review, independent checks and coordinator ordinary-browser observations; same model family, not a human or Claude assessment",
          "assessed_at": "2026-11-14T18:58:26.756115+00:00",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 4,
            "B": 4
          },
          "reasons": {
            "C": "Independently recomputed80/68 packages,70/82 remaining allowance,8 Gia-days and reportsD15/D11 with nonoverlapping schedules. Checked48 combinations for cost/capacity/blocking rules. Audit correctly distinguishes33.3% shorter elapsed time, equal24 measured crew-weeks and active2.4 versus2.0MW/crew-week; K remains impossible at D10 against D20 approval.",
            "D": "A concrete staged80 package with30 separable D1 investigation and50 D5 balance;68 alternative explains cost/coverage/report-date tradeoff. Named owners, bad/late evidence hold/partial-release conditions, board-decline route and operational controls make it actionable.",
            "E": "Adverse histories, competing explanations and missing counterfactuals are retained. Sampling sufficiency and design are Gia decisions, not invented certification; road action is justified from current vulnerability without invented probabilities or savings.",
            "F": "Selected and rejected lessons link evidence, mechanism, boundaries, falsifiers, baseline changes, budget, engineering capacity, authority and deadline consequences. Further redesign/remediation remains unfunded/unproved rather than hidden.",
            "M": "Three-part recommendation and dated authority position lead; package controls, audit, lesson cards and owner/trigger register provide an easy route to detail. Historical costs, current offers and proposal/approval labels remain distinct.",
            "B": "Coordinator observed added P/K143k and16/8 failure, delayed G16D19 hold, no-investigation supportive outcome still held and reload. Independent JSDOM covered48 packages, delayed/invalid dates, conditional ground outcomes, reset and all anchors."
          },
          "material_failures": [],
          "findings": [],
          "verification_gaps": [
            "This is a separate same-family AI assessment, not practitioner checking, Claude validation or real-world authority.",
            "Native browser observations were supplied by the coordinator; the reviewer independently ran JSDOM and source-derived checks, not a separate native-browser session.",
            "No claim of exhaustive browser-engine, keyboard-accessibility or responsive-layout coverage."
          ],
          "disposition": "verification only",
          "independent_checks": [
            "independent-results.json:48 ground/optional-measure combinations checked for totals and blocking conditions.",
            "Recommended G16fieldD2–11/analysisD12–15 and I D6–9; alternativeG8fieldD2–7/analysisD8–11 and I D2–5.",
            "G16D5approval producesD19 report andHOLD; no-ground supportive outcome stillHOLD; invalid date flagged."
          ],
          "artifact_sha256": "7f2f504654d160affa77a72bab6321441de8ed24fb616fa3605d121bb32c6d04",
          "status": "provisional AI assessment",
          "rationale": "AI review disposition: verification only",
          "served_artifact_sha256": "3b7e646442e4a3a88db90f02746a52b7c56b7fdde43717d12cf6649350f262fb",
          "review_evidence_id": "essai03-separate-review-15-20"
        },
        "current": {
          "case_id": "BPB2-19",
          "reviewer": "gpt-6-astra · separate assessment agent",
          "method": "Separate source/logic review, independent checks and coordinator ordinary-browser observations; same model family, not a human or Claude assessment; current case analysis is unchanged from the retained first answer",
          "assessed_at": "2026-11-14T18:58:26.756115+00:00",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 4,
            "B": 4
          },
          "reasons": {
            "C": "Independently recomputed80/68 packages,70/82 remaining allowance,8 Gia-days and reportsD15/D11 with nonoverlapping schedules. Checked48 combinations for cost/capacity/blocking rules. Audit correctly distinguishes33.3% shorter elapsed time, equal24 measured crew-weeks and active2.4 versus2.0MW/crew-week; K remains impossible at D10 against D20 approval.",
            "D": "A concrete staged80 package with30 separable D1 investigation and50 D5 balance;68 alternative explains cost/coverage/report-date tradeoff. Named owners, bad/late evidence hold/partial-release conditions, board-decline route and operational controls make it actionable.",
            "E": "Adverse histories, competing explanations and missing counterfactuals are retained. Sampling sufficiency and design are Gia decisions, not invented certification; road action is justified from current vulnerability without invented probabilities or savings.",
            "F": "Selected and rejected lessons link evidence, mechanism, boundaries, falsifiers, baseline changes, budget, engineering capacity, authority and deadline consequences. Further redesign/remediation remains unfunded/unproved rather than hidden.",
            "M": "Three-part recommendation and dated authority position lead; package controls, audit, lesson cards and owner/trigger register provide an easy route to detail. Historical costs, current offers and proposal/approval labels remain distinct.",
            "B": "Coordinator observed added P/K143k and16/8 failure, delayed G16D19 hold, no-investigation supportive outcome still held and reload. Independent JSDOM covered48 packages, delayed/invalid dates, conditional ground outcomes, reset and all anchors."
          },
          "material_failures": [],
          "findings": [],
          "verification_gaps": [
            "This is a separate same-family AI assessment, not practitioner checking, Claude validation or real-world authority.",
            "Native browser observations were supplied by the coordinator; the reviewer independently ran JSDOM and source-derived checks, not a separate native-browser session.",
            "No claim of exhaustive browser-engine, keyboard-accessibility or responsive-layout coverage."
          ],
          "disposition": "verification only",
          "independent_checks": [
            "independent-results.json:48 ground/optional-measure combinations checked for totals and blocking conditions.",
            "Recommended G16fieldD2–11/analysisD12–15 and I D6–9; alternativeG8fieldD2–7/analysisD8–11 and I D2–5.",
            "G16D5approval producesD19 report andHOLD; no-ground supportive outcome stillHOLD; invalid date flagged."
          ],
          "artifact_sha256": "7f2f504654d160affa77a72bab6321441de8ed24fb616fa3605d121bb32c6d04",
          "status": "provisional AI assessment",
          "rationale": "AI review disposition: verification only",
          "served_artifact_sha256": "a269176930cd240688bcf0f9a8a9abe42fd8ba896a87f8c8862bd6f7177c0e62",
          "review_evidence_id": "essai03-separate-review-15-20"
        }
      },
      "status": "first attempt retained"
    },
    {
      "id": "BPB2-20",
      "self": {
        "initial": {
          "reviewer": "gpt-6-astra · answering agent",
          "method": "Unverified self-assessment before feedback; common assessment/source packaging added afterward",
          "assessed_at": "2026-11-14T17:54:55Z",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 3,
            "B": 3
          },
          "reasons": {
            "C": "Reconciles £10,000 signed base contracts, £3,200 paid, £6,800 outstanding, £4,000 essentials and £0 free base cash. Preserves H8, Saturday, V1 completion, V2 revision and exact interval/lead deficits. All eight staffing combinations checked.",
            "D": "Leads with a viable proposed H8 amendment, exact motion, coordinated signing/direct-payment requirements, earlier Monday cutoff, failure/extension route and owned plan through 20 July.",
            "E": "Source-linked material claims, full original records, fixed actual unsigned banner, labelled forecasts/proposals, unresolved £6,000 refusal treatment and no invented permission, confirmations or signatures.",
            "F": "Provides all requested sections, five status filters with recursive prerequisite retention, P4 corrections, cash bridge, role/interval coverage, document control, recovery/reset and supporting evidence.",
            "M": "Readable practitioner memo with lead recommendation and navigable sections; long register and evidence make the page substantial. Narrow-screen table behavior is designed but not visually verified.",
            "B": "Eleven JSDOM assertion groups passed including all controls, source links, dependency closure, reset and fresh reload. No script errors. Actual browser rendering and visual interaction remain for coordinator verification."
          },
          "material_failures": [],
          "rationale": "Self-assessment only.",
          "artifact_sha256": "8591ec89140854ce595c1c66d7fcca9a7332ad35c63ad9ed093f2570e35c27e5",
          "verification_gaps": []
        },
        "current": {
          "reviewer": "gpt-6-astra · answering agent",
          "method": "Unverified self-assessment before feedback; common assessment/source packaging added afterward; current case analysis unchanged",
          "assessed_at": "2026-11-14T17:54:55Z",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 3,
            "B": 3
          },
          "reasons": {
            "C": "Reconciles £10,000 signed base contracts, £3,200 paid, £6,800 outstanding, £4,000 essentials and £0 free base cash. Preserves H8, Saturday, V1 completion, V2 revision and exact interval/lead deficits. All eight staffing combinations checked.",
            "D": "Leads with a viable proposed H8 amendment, exact motion, coordinated signing/direct-payment requirements, earlier Monday cutoff, failure/extension route and owned plan through 20 July.",
            "E": "Source-linked material claims, full original records, fixed actual unsigned banner, labelled forecasts/proposals, unresolved £6,000 refusal treatment and no invented permission, confirmations or signatures.",
            "F": "Provides all requested sections, five status filters with recursive prerequisite retention, P4 corrections, cash bridge, role/interval coverage, document control, recovery/reset and supporting evidence.",
            "M": "Readable practitioner memo with lead recommendation and navigable sections; long register and evidence make the page substantial. Narrow-screen table behavior is designed but not visually verified.",
            "B": "Eleven JSDOM assertion groups passed including all controls, source links, dependency closure, reset and fresh reload. No script errors. Actual browser rendering and visual interaction remain for coordinator verification."
          },
          "material_failures": [],
          "rationale": "Self-assessment only.",
          "artifact_sha256": "8591ec89140854ce595c1c66d7fcca9a7332ad35c63ad9ed093f2570e35c27e5",
          "verification_gaps": []
        }
      },
      "checked": {
        "initial": {
          "case_id": "BPB2-20",
          "reviewer": "gpt-6-astra · separate assessment agent",
          "method": "Separate source/logic review, independent checks and coordinator ordinary-browser observations; same model family, not a human or Claude assessment",
          "assessed_at": "2026-11-14T18:58:26.756115+00:00",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 4,
            "B": 4
          },
          "reasons": {
            "C": "Independently reconciled10000 signed,3200 paid,6800 payable plus4000 essentials against10800 bank, free0; sponsor direct2000 is inside8000/5000 and full route leaves3000. Derived interval counts6/7/8/7 and leads2/1/2/2 across all8 roster variants. H8 authority and fixed deadline sequence correctly interpreted.",
            "D": "Gives concrete proposed H9 wording, two-affirmative-including-Sam adoption, paired contracts/direct deposit before Monday10, essential quotes before noon, targeted staffing recovery and extension/lapse conditions. No signature or council permission is presumed.",
            "E": "Preserves valid H8, completedV1 and original action deadlines; superseded P4/draft Sunday statements are corrected explicitly. Sponsor refusal terms for remaining6000 remain unknown with a conservative exclusion clearly labelled, not an invented refund liability.",
            "F": "Current status, dependency-aware register, full P4 corrections, base/economic/cash bridges, role/interval staffing, exact board proposal and dated route through20-Jul cover the whole requested decision record.",
            "M": "Opening makes Saturday/unsigned/no-free-cash position unmistakable. Status filters retain prerequisites; fixed actual position remains visible beside hypothetical finance/staffing views; detailed board language and evidence remain easy to reach.",
            "B": "Coordinator observed conditional filter dependencies, full-route3000, Kai lead gap, Hana closure and reset. Independent JSDOM checked all5 filters for dependency closure, all8 staffing combinations, each cash/refusal/receipt view, reset and local anchors."
          },
          "material_failures": [],
          "findings": [
            {
              "severity": "observation",
              "description": "The offer leaves refusal treatment of the remaining6000 sponsorship unspecified. The response identifies this and makes paired-document clarification an explicit safeguard without claiming a refund term already exists.",
              "evidence": "20-G and answer Accounting assumptions and the refusal boundary / proposed H9.",
              "proposed_repair": "None for this synthetic decision brief. Actual execution would require the stated terms to be agreed and documented."
            }
          ],
          "verification_gaps": [
            "This is a separate same-family AI assessment, not practitioner checking, Claude validation or real-world authority.",
            "Native browser observations were supplied by the coordinator; the reviewer independently ran JSDOM and source-derived checks, not a separate native-browser session.",
            "No claim of exhaustive browser-engine, keyboard-accessibility or responsive-layout coverage."
          ],
          "disposition": "verification only",
          "independent_checks": [
            "independent-results.json:all8 hypothetical roster combinations derived from person intervals/roles and compared to rendered counts.",
            "All5 required status filters preserve recursive prerequisites and rendered evidence targets.",
            "Actual/paired/permitted/refused cash views checked with and without6000 received; only permitted full-sponsorship economics yields3000 freely available."
          ],
          "artifact_sha256": "8591ec89140854ce595c1c66d7fcca9a7332ad35c63ad9ed093f2570e35c27e5",
          "status": "provisional AI assessment",
          "rationale": "AI review disposition: verification only",
          "served_artifact_sha256": "5800ed688595830d3eab6b14ad42e5837572f37818e0364285b9b21570d69d90",
          "review_evidence_id": "essai03-separate-review-15-20"
        },
        "current": {
          "case_id": "BPB2-20",
          "reviewer": "gpt-6-astra · separate assessment agent",
          "method": "Separate source/logic review, independent checks and coordinator ordinary-browser observations; same model family, not a human or Claude assessment; current case analysis is unchanged from the retained first answer",
          "assessed_at": "2026-11-14T18:58:26.756115+00:00",
          "dimensions": {
            "C": 4,
            "D": 4,
            "E": 4,
            "F": 4,
            "M": 4,
            "B": 4
          },
          "reasons": {
            "C": "Independently reconciled10000 signed,3200 paid,6800 payable plus4000 essentials against10800 bank, free0; sponsor direct2000 is inside8000/5000 and full route leaves3000. Derived interval counts6/7/8/7 and leads2/1/2/2 across all8 roster variants. H8 authority and fixed deadline sequence correctly interpreted.",
            "D": "Gives concrete proposed H9 wording, two-affirmative-including-Sam adoption, paired contracts/direct deposit before Monday10, essential quotes before noon, targeted staffing recovery and extension/lapse conditions. No signature or council permission is presumed.",
            "E": "Preserves valid H8, completedV1 and original action deadlines; superseded P4/draft Sunday statements are corrected explicitly. Sponsor refusal terms for remaining6000 remain unknown with a conservative exclusion clearly labelled, not an invented refund liability.",
            "F": "Current status, dependency-aware register, full P4 corrections, base/economic/cash bridges, role/interval staffing, exact board proposal and dated route through20-Jul cover the whole requested decision record.",
            "M": "Opening makes Saturday/unsigned/no-free-cash position unmistakable. Status filters retain prerequisites; fixed actual position remains visible beside hypothetical finance/staffing views; detailed board language and evidence remain easy to reach.",
            "B": "Coordinator observed conditional filter dependencies, full-route3000, Kai lead gap, Hana closure and reset. Independent JSDOM checked all5 filters for dependency closure, all8 staffing combinations, each cash/refusal/receipt view, reset and local anchors."
          },
          "material_failures": [],
          "findings": [
            {
              "severity": "observation",
              "description": "The offer leaves refusal treatment of the remaining6000 sponsorship unspecified. The response identifies this and makes paired-document clarification an explicit safeguard without claiming a refund term already exists.",
              "evidence": "20-G and answer Accounting assumptions and the refusal boundary / proposed H9.",
              "proposed_repair": "None for this synthetic decision brief. Actual execution would require the stated terms to be agreed and documented."
            }
          ],
          "verification_gaps": [
            "This is a separate same-family AI assessment, not practitioner checking, Claude validation or real-world authority.",
            "Native browser observations were supplied by the coordinator; the reviewer independently ran JSDOM and source-derived checks, not a separate native-browser session.",
            "No claim of exhaustive browser-engine, keyboard-accessibility or responsive-layout coverage."
          ],
          "disposition": "verification only",
          "independent_checks": [
            "independent-results.json:all8 hypothetical roster combinations derived from person intervals/roles and compared to rendered counts.",
            "All5 required status filters preserve recursive prerequisites and rendered evidence targets.",
            "Actual/paired/permitted/refused cash views checked with and without6000 received; only permitted full-sponsorship economics yields3000 freely available."
          ],
          "artifact_sha256": "8591ec89140854ce595c1c66d7fcca9a7332ad35c63ad9ed093f2570e35c27e5",
          "status": "provisional AI assessment",
          "rationale": "AI review disposition: verification only",
          "served_artifact_sha256": "fc5f6f8262cf6e8ba5c8f37312532d6895beb727f478212834a84b3d93fcce69",
          "review_evidence_id": "essai03-separate-review-15-20"
        }
      },
      "status": "first attempt retained"
    }
  ],
  "model": "gpt-6-astra",
  "reasoning": "ultra",
  "status": "completed-development-run"
}
